Cifar-10 Image Generation GAN

Name: Soh Hong Yu

Admin Number: P2100775

Class: DAAA/FT/2B/01

Module Code: ST1504 Deep Learning

---

References (In Harvard format):

1. Goodfellow, I.J. et al. (2014) Generative Adversarial Networks, arXiv.org.
   Available at: https://arxiv.org/abs/1406.2661 (Accessed: January 21, 2023).
2. Cox, S. (2021) The overlooked technique of image averaging, Photography Life.
   Available at: https://photographylife.com/image-averaging-technique (Accessed: November 19, 2022).
3. Brownlee, J. (2019) How to evaluate generative Adversarial Networks, MachineLearningMastery.com.
   Available at: https://machinelearningmastery.com/how-to-evaluate-generative-adversarial-networks/ (Accessed: January 21, 2023).
4. Hui, J. (2020) Gan - ways to improve gan performance, Medium. Towards Data Science.
   Available at: https://towardsdatascience.com/gan-ways-to-improve-gan-performance-acf37f9f59b (Accessed: January 21, 2023).
5. Briggs, J. (2020) Beating the gan game, Medium. Towards Data Science.
   Available at: https://towardsdatascience.com/beating-the-gan-game-afbcce0a20be (Accessed: January 21, 2023).
6. Silva, T.S. (2017) A short introduction to generative Adversarial Networks, A Short Introduction to Generative Adversarial Networks - Thalles' blog.
   Available at: https://sthalles.github.io/intro-to-gans/ (Accessed: January 21, 2023).
7. A. Radford, L. Metz, en S. Chintala, "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks", arXiv [cs.LG]. 2016.
8. T. Miyato, T. Kataoka, M. Koyama, en Y. Yoshida, "Spectral Normalization for Generative Adversarial Networks", arXiv [cs.LG]. 2018.
9. Hui, J. (2021) Gan - spectral normalization, Medium. Medium.
   Available at: https://jonathan-hui.medium.com/gan-spectral-normalization-893b6a4e8f53 (Accessed: January 21, 2023).
10. Salimans, T. and Kingma, D.P. (2016) Weight normalization: A simple reparameterization to accelerate training of deep neural networks, arXiv.org.
    Available at: https://arxiv.org/abs/1602.07868 (Accessed: January 21, 2023).
11. Tensorflow (2023) Customize what happens in model.fit  :   Tensorflow Core, TensorFlow.
    Available at: https://www.tensorflow.org/guide/keras/customizing_what_happens_in_fit#wrapping_up_an_end-to-end_gan_example (Accessed: January 21, 2023).
12. Google (2023) Common problems  |  machine learning  |  google developers, Google. Google.
    Available at: https://developers.google.com/machine-learning/gan/problems (Accessed: January 21, 2023).
13. Hui, J. (2019) Gan - why it is so hard to train generative adversarial networks!, Medium. Medium.
    Available at: https://jonathan-hui.medium.com/gan-why-it-is-so-hard-to-train-generative-advisory-networks-819a86b3750b (Accessed: January 21, 2023).
14. Lazarou, C. (2020) Why do gans need so much noise?, Medium. Towards Data Science.
    Available at: https://towardsdatascience.com/why-do-gans-need-so-much-noise-1eae6c0fb177 (Accessed: January 21, 2023).
15. Brownlee, J. (2019) How to calculate the KL divergence for Machine Learning, MachineLearningMastery.com.
    Available at: https://machinelearningmastery.com/divergence-between-probability-distributions/ (Accessed: January 21, 2023).
16. Agrawal, T. (2022) Gans failure modes: How to identify and monitor them, neptune.ai.
    Available at: https://neptune.ai/blog/gan-failure-modes (Accessed: January 21, 2023).
17. Brownlee, J. (2019) How to implement the Frechet Inception Distance (FID) for evaluating Gans, MachineLearningMastery.com.
    Available at: https://machinelearningmastery.com/how-to-implement-the-frechet-inception-distance-fid-from-scratch/ (Accessed: January 21, 2023).
18. Papers with code (2023) Papers with code - leaky relu explained, Explained | Papers With Code.
    Available at: https://paperswithcode.com/method/leaky-relu (Accessed: January 21, 2023).
19. Papers with code (2023) Papers with code - prelu explained, PReLU Explained | Papers With Code.
    Available at: https://paperswithcode.com/method/prelu (Accessed: January 21, 2023).
20. Mishra, D. (2021) Transposed convolution demystified, Medium. Towards Data Science.
    Available at: https://towardsdatascience.com/transposed-convolution-demystified-84ca81b4baba (Accessed: January 21, 2023).
21. Thakur, A. (2021) How to evaluate Gans using Frechet Inception Distance (FID), W&B.
    Available at: https://wandb.ai/ayush-thakur/gan-evaluation/reports/How-to-Evaluate-GANs-using-Frechet-Inception-Distance-FID---Vmlldzo0MTAxOTI (Accessed: January 21, 2023).
22. Viswanath, P. (2017) 12.1 generative adversarial network - courses.engr.illinois.edu, ECE598: Representation Learning: Algorithms and Models.
    Available at: https://courses.engr.illinois.edu/ece598pv/fa2017/Lecture12_GAN_AmirT.pdf (Accessed: January 20, 2023).
23. Mescheder, L., Geiger, A. and Nowozin, S. (2018) Which training methods for gans do actually converge?, arXiv.org.
    Available at: https://arxiv.org/abs/1801.04406 (Accessed: January 21, 2023).
24. Kodali, N. et al. (2017) On convergence and stability of gans, arXiv.org.
    Available at: https://arxiv.org/abs/1705.07215 (Accessed: January 21, 2023).
25. Papers with code (2023) Papers with code - CIFAR-10 benchmark (image generation), The latest in Machine Learning.
    Available at: https://paperswithcode.com/sota/image-generation-on-cifar-10 (Accessed: January 29, 2023).
26. Jung, S. and Keuper, M. (2021) Internalized biases in Fréchet inception distance, OpenReview.
    Available at: https://openreview.net/forum?id=mLG96UpmbYz (Accessed: January 29, 2023).

Project Objective

Implement a suitable GAN architecture to the problem. Creating 1000 new images by training the model on the cifar-10 dataset.

Background Information

In the world where data is very important, there are certain information that is sensitive or private. Using GAN allow us to generate new and unseen images which can be used for training of other models etc.

CIFAR-10 dataset (Canadian Institute for Advanced Research) is a collection of images that are commonly used to train and benchmark image classification algorithms. It is a subset of 80 million tiny images dataset and consists of 60,000 32 x 32 color images in 10 different classes. There are 50000 training images and 10000 test images. It was collected by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. [1]

Initialising Libraries and Variables

import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import keras.backend as K
from tensorflow import keras
from keras.layers import Reshape, Conv2DTranspose, PReLU
from keras.utils import to_categorical, plot_model
from keras.layers import Concatenate
import numpy as np
from keras.applications.inception_v3 import InceptionV3, preprocess_input
from tensorflow.image import resize
from scipy.linalg import sqrtm
import math
from tqdm.notebook import tqdm
import tensorflow as tf
from IPython.display import clear_output, HTML
import imageio
import glob
from keras.layers import AveragePooling2D, ZeroPadding2D, BatchNormalization, Activation, MaxPool2D, Add
from keras.layers import Normalization, Dense, Conv2D, Dropout, BatchNormalization, ReLU
from keras.models import Sequential, Model
from keras import Input
from keras.optimizers import *
from keras.callbacks import EarlyStopping
import visualkeras
from keras.initializers import RandomNormal
from tensorflow_addons.layers import SpectralNormalization
from keras.layers import LeakyReLU, GlobalMaxPooling2D, GlobalAveragePooling2D

# Fix random seed for reproducibility
seed = 88
np.random.seed(seed)

Checking GPU

# Check if Cuda GPU is available
tf.config.list_physical_devices('GPU')

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Loading Datasets

df = tf.keras.datasets.cifar10.load_data()

(x_train, y_train), (x_test, y_test) = df

Exploratory Data Analysis

We will begin by conducting an exploratory data analysis of the data, to gain a better understanding of the characteristics of the dataset.

x_train: uint8 NumPy array of coloured image data with shapes (50000, 32, 32, 3), containing the training data. Pixel values range from 0 to 255.

y_train: uint8 NumPy array of labels (integers in range 0-9) with shape (50000, 1) for the training data.

x_test: uint8 NumPy array of coloured image data with shapes (10000, 32, 32, 3), containing the test data. Pixel values range from 0 to 255.

y_test: uint8 NumPy array of labels (integers in range 0-9) with shape (10000, 1) for the test data.

There are 10 different type of labels in the dataset. From the dataset, each value represent an item. The following list is the description of each value.

Item Labels

• 0 : Airplane
• 1 : Automobile
• 2 : Bird
• 3 : Cat
• 4 : Deer
• 5 : Dog
• 6 : Frog
• 7 : Horse
• 8 : Ship
• 9 : Truck

class_labels = {
    0: 'Airplane',
    1: 'Automobile',
    2: 'Bird',
    3: 'Cat',
    4: 'Deer',
    5: 'Dog',
    6: 'Frog',
    7: 'Horse',
    8: 'Ship',
    9: 'Truck'
}

NUM_CLASS = 10

Each image is a 32x32 image as well as 3 color channel [RGB] (coloured image). Therefore, we can set the IMG_SIZE as a tuple (32, 32, 3)

IMG_SIZE = (32, 32, 3)

Visualising the Dataset

Let's look at what the images look like.

fig, ax = plt.subplots(10, 10, figsize=(30, 30))
for i in range(10):
    images = x_train[np.squeeze(y_train == i)]
    random_index = np.random.choice(images.shape[0], 10, replace=False)
    images = images[random_index]
    label = class_labels[i]
    for j in range(10):
        subplot = ax[i, j]
        subplot.axis("off")
        subplot.imshow(images[j])
        subplot.set_title(label)

plt.show()

Observation

We notes that the images are correctly labelled. However, there are some concerns which the AI might use to help generate these images. For example, the cat image has text around the cat. This might be an issue as the AI might mistaken the cat and try to generate images of text inside which can affect our desired results. As well as the different species of bird like an ostrich and sparrow. This might be an issue as the AI will not be able to clearly differentiate the different species and generate an amalgamation. Another issue with the data is the zoom of the images. The images have different sizing like for the dogs and cat. Sometimes the images are just the face of the animal or sometimes the images are a zoomed out version and the entire body of the animal can be seen.

Class Distribution

When training a machine learning model, it is always important to check the distribution of the different classes in the dataset. This will inform us which metrics is the best to use and if anything is needed to balance the classes.

labels, counts = np.unique(y_train, return_counts=True)
for label, count in zip(labels, counts):
    print(f"{class_labels[label]}: {count}")

Airplane: 5000
Automobile: 5000
Bird: 5000
Cat: 5000
Deer: 5000
Dog: 5000
Frog: 5000
Horse: 5000
Ship: 5000
Truck: 5000

plt.figure(figsize=(20, 15))
plt.barh(labels, counts, tick_label=list(class_labels.values()))
plt.show()

Observations

As we can see from the bar graph, the distribution of the images is even. And there will be no bias towards one particular class

Image Pixel Distribution

We need to know the pixel intensity and know the distribution of the pixels

print("Max: ", np.max(x_train))
print("Min: ", np.min(x_train))

Max:  255
Min:  0

As expected, our pixels have values between 0 and 255.

mean, std = np.mean(x_train, axis=(0, 1, 2)), np.std(x_train, axis=(0, 1, 2))
print("Mean:", mean)
print("std:", std)

Mean: [125.30691805 122.95039414 113.86538318]
std: [62.99321928 62.08870764 66.70489964]

Image Averaging

Image Averaging involves stacking multiple photos on top of each other and averaging them together. The main purpose is to see the noise of the image adn therefore reducing it.

plt.imshow(np.mean(x_train, axis=0) / 255)
plt.show()

Observation

We cannot see a single thing from the image. This is likely due to the color of the images overlaying each other giving this blur effect.

fig, ax = plt.subplots(2, 5, figsize=(25, 10))

for idx, subplot in enumerate(ax.ravel()):
    avg_image = np.mean(x_train[np.squeeze(y_train == idx)], axis=0) / 255
    subplot.imshow(avg_image)
    subplot.set_title(f"{class_labels[idx]}")
    subplot.axis("off")

Observations

Although the average images is blurry, we can make out the images of an automobile, horse and truck etc. It is more difficult to make out the average image for the other classes. This suggests that for these similar class, the generated images might be able to learn more robust and representative features of the training data which means the images that will be generated will be very good.

Data Preprocessing

Before modelling, its is important to perform data preprocessing

One Hot Encoding

As they are, the current labels are encoded from 0-10, we will one hot encode the labels.

y_train = to_categorical(y_train)
y_test = to_categorical(y_test)

print(y_train[0])
print("Label:", tf.argmax(y_train[0]))

[0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]
Label: tf.Tensor(6, shape=(), dtype=int64)
Label: tf.Tensor(6, shape=(), dtype=int64)

Normalizing the inputs

Image normalisation is done to the dataset.

Normalising the inputs means that we will calculate the mean and standard deviation of the training set, and then apply the formula below.

$$X_{} = \frac{X_{} - μ_{}}{σ_{}}$$

Pixel values of each pixel are on similar scale, therefore normalisation can be used. This helps to optimize the algorithm to better converge during gradient descent.

pre_processing_v1 = Normalization()
pre_processing_v1.adapt(x_train)

fig, ax = plt.subplots(ncols=2)

ax[0].imshow(x_train[0])
ax[0].set_title('Before Preprocessing')
ax[1].imshow(tf.squeeze(pre_processing_v1(x_train[:1, :, :])))
ax[1].set_title('After Preprocessing')
plt.show()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

x_train = x_train.astype('float32')
x_train /= (255/2)
x_train -= 1

Building Models

In a generative adversarial network [GAN] there will 2 networks in play. A generator and a discriminator.

Generator: Produce synthetic data samples similar to the training data Discriminator: Distinguish samples from real dataset to the generator

Both networks plays a game of MinMax where the generator wants to minimize the difference image to make it look real and the discriminator wants to maximize the ability to spot the real from the fakes.

$$\min _{G} \max _{D} V(D, G)=\mathbb{E}_{\boldsymbol{x} \sim p_{\text {data }}(\boldsymbol{x})}[\log D(\boldsymbol{x})]+\mathbb{E}_{\boldsymbol{z} \sim p_{\boldsymbol{z}}(\boldsymbol{z})}[\log (1-D(G(\boldsymbol{z})))]$$

We will be building a few deep learning models to generate and identify real or fake images.

Architecture List:

1. DCGAN (Baseline)
2. cGAN
3. ACGAN

However before we create the generator, we should look through the different models that can be used to make our discriminator too.

As the discriminator is only predicting real image vs fake image, the discriminator will be a CNN model.

GAN Model Evaluation Methods

There are several methods to evaluate the GAN. However, there are some issues with evaluating a GAN generator model.

1. There is no labels for GAN, the main focus is the quality of the synthetic data.
2. Different models have different objective functions

There are a few methods:

1. Manual GAN Evaluation
   • The simplest evaluation method is through visual evaluation using our eyes.
   • However, there are some issues:
     1. This method is very subjective and have some bias by the reviewer
     2. As the Cifar-10 images are only (32, 32, 3), it is hard to differentiate the classes and it is very small which is unclear
     3. If there is a large quantity of images generated, there is no way to go through each and every image to evaluate it
2. Qualitative GAN Evaluation
   • By identifying non numerical data in the images, we are able to use human subject evaluation or evaluation via comparison. Some techniques include using:
     1. Nearest Neighbors [Detect overfitting, generated samples are shown next to nearest neighbour in the training set]
     2. Rapid Scene Categorization [Get participants to help classify real and fake data]
3. Quantitative GAN Evaluation
   • Evaluating GAN on quantitative means refers to using a specific calculation of specific numerical scores, used to summarize the general quality of the images.
   • There are a few metrics:
     1. KL Divergence (aka relative entropy) [Measures the distance between two statistical probability]
     2. Fréchet Inception Distance [compares distribution of generated images]

For simplicity, we will be only evaluating the GAN models with manual and quantitative evaluation

Fréchet inception distance (FID)

Fréchet inception distance (FID) is a metric that calculates the distance between feature vectors calculated for real and generated images. Basically it shows how closely related the generated image and real image are to one another. The lower the score the more correlated the images are [Might suggest higher quality images].

$$FID = ||\mu - \mu_w||^2_2 + tr(C + C_w - 2(C C_w)^\frac{1}{2})$$

We also need to take note that FID score may not be the most accurate as FID score requires a large dataset to be very precise as FID uses covariance to calculate the score. Further more FID score is biases as the InceptionV3 model used to extract features are corrupted by color, intensity, saturation, and horizontal translations. Another issue with the FID score is we are unable to tell which classes is generated better.

However, as FID is one of the most common GAN evaluation metrics, we will ignore this issues with FID score and continue using this as our comparing metrics

class GAN_FID:
    def __init__(self, batch_size, noise, sample_size, buffer_size, labels=False):
        # setting Hyperparameters
        self.BATCH_SIZE = batch_size
        self.noise = noise
        self.SAMPLE_SIZE = sample_size
        self.BUFFER_SIZE = buffer_size
        self.labels = labels

        # setting Constants
        self.INCEPTION_SHAPE = (299, 299, 3)
        self.INCEPTION = InceptionV3(
            include_top=False, pooling='avg', input_shape=self.INCEPTION_SHAPE)
        self.AUTO = tf.data.AUTOTUNE

    # method to set generator and training data
    def fit(self, generator, train_data):
        # setting generative model and original data used for training
        self.GENERATOR = generator
        self.train_data = train_data

        # Preparing Real Images
        trainloader = tf.data.Dataset.from_tensor_slices((self.train_data))
        trainloader = (
            trainloader
            .shuffle(self.BUFFER_SIZE)
            .map(self.__resize_and_preprocess, num_parallel_calls=self.AUTO)
            .batch(self.BATCH_SIZE, num_parallel_calls=self.AUTO)
            .prefetch(self.AUTO)
        )
        self.trainloader = trainloader

        if self.labels:
            # Generate and prepare Synthetic Images (Fake)
            rand_labels = np.random.randint(
                low=0, high=10, size=self.SAMPLE_SIZE)
            rand_labels = to_categorical(rand_labels)
        noise = tf.random.normal([self.SAMPLE_SIZE, self.noise])
        if self.labels:
            generated_images = self.GENERATOR([noise, rand_labels])
        else:
            generated_images = self.GENERATOR(noise)
        genloader = tf.data.Dataset.from_tensor_slices(generated_images)
        genloader = (
            genloader
            .map(self.__resize_and_preprocess, num_parallel_calls=self.AUTO)
            .batch(self.BATCH_SIZE, num_parallel_calls=self.AUTO)
            .prefetch(self.AUTO)
        )
        self.genloader = genloader

        # prepare embeddings
        count = math.ceil(self.SAMPLE_SIZE/self.BATCH_SIZE)

        # compute embeddings for real images
        print("Computing Real Image Embeddings")
        self.real_image_embeddings = self.__compute_embeddings(
            self.trainloader, count)

        # compute embeddings for generated images
        print("Computing Generated Image Embeddings")
        self.generated_image_embeddings = self.__compute_embeddings(
            self.genloader, count)
        assert self.real_image_embeddings.shape == self.generated_image_embeddings.shape, "Embeddings are not of the same size"
        print("Computed Embeddings\tReal Images Embedding Shape: {}\tGenerated Images Embedding Shape: {}".format(
            self.real_image_embeddings.shape,
            self.generated_image_embeddings.shape
        ))

    # method to produce evaluation results
    @tf.autograph.experimental.do_not_convert
    def evaluate(self):
        # calculate Frechet Inception Distance
        fid = self.__calculate_fid(
            self.real_image_embeddings, self.generated_image_embeddings)
        print('The computed FID score is:', fid)

        return fid

    # method to generate embeddings from inception model
    def __compute_embeddings(self, dataloader, count):
        image_embeddings = []
        for _ in tqdm(range(count)):
            images = next(iter(dataloader))
            embeddings = self.INCEPTION.predict(images)
            image_embeddings.extend(embeddings)
        return np.array(image_embeddings)

    # STATIC METHODS: these methods knows nothing about the class
    # static method to prepare the data before computing Inception Embeddings
    @staticmethod
    def __resize_and_preprocess(image):
        # image *= 255.0 # original image are scaled to [0, 1], scaling back to [0, 255]
        image -= -1
        image /= (1 - (-1))
        image *= 255.

        # .preprocess_input() expects an image of scale [0, 255]
        image = preprocess_input(image)
        # inception model expects an image of shape (None, 299, 299, 3)
        image = tf.image.resize(image, (299, 299), method='nearest')
        return image

    # static method to calculate frechet inception distance based on embeddings
    @staticmethod
    def __calculate_fid(real_embeddings, generated_embeddings):
        # calculate mean and covariance statistics
        mu1, sigma1 = real_embeddings.mean(
            axis=0), np.cov(real_embeddings, rowvar=False)
        mu2, sigma2 = generated_embeddings.mean(
            axis=0), np.cov(generated_embeddings, rowvar=False)
        # calculate sum squared difference between means
        ssdiff = np.sum((mu1 - mu2)**2.0)
        # calculate sqrt of product between cov
        covmean = sqrtm(sigma1.dot(sigma2))
        # check and correct imaginary numbers from sqrt
        if np.iscomplexobj(covmean):
            covmean = covmean.real
        # calculate score
        fid = ssdiff + np.trace(sigma1 + sigma2 - 2.0 * covmean)
        return fid

# Hyperparameters and Constants
BATCH_SIZE = 64
AUTO = tf.data.AUTOTUNE
EPOCHS = 100
BUFFER_SIZE = 1024
NOISE = 128

DCGAN (Baseline)

For a GAN architecture, it consist of 2 different models.

1. Discriminator
2. Generator

Both Discriminator and Generator will play a min max game. The generator wants to minimize and ensure that the image generated looks real. The discriminator wants to maximize and ensure that it can distinguish samples from real and fake and wants to spot the fake.

DCGAN Discriminator

Basically, the discriminator is a binary classifier. For the DCGAN model, we will make a typical CNN network. However, unlike a typical CNN network, we will be modifying it such that DCGAN model will be improved.

1. Pooling Layers in a classic CNN are used to down sample activation features maps. However, this is not recommended for GAN as information will be loss. Therefore, we will be using a 2 x 2 Stride Convolution
2. Due to the vanishing gradient problem, the activation function for a GAN is usually a LeakyReLu. It allows gradients to flow easier through the architecture. This allows the function to compute the greatest value between the features and a small factor.
3. As sparse gradients will occur and diminish over time [random spikes in the loss during training], there will be many numerical operations performed on them. To fix this, we will add BatchNormalization to the network so that it can reduce covariate shifts. BatchNormalization will be used before non-Gaussian activations like LeakyReLU.
4. One issue of implementing GAN is the unstable training issue. The issue happens when the discriminator becomes more powerful and overwhelming the generator. We will normalise the weights and controls the Lipschitz constant of the discriminator to mitigate the exploding gradient problem and the mode collapse problem. [Spectral Normalization]
5. He and Xavier weights initialization are unsuitable for training as it makes training more unstable due to the larger distribution of weights. To reduce the instability of training, we will be using Gaussian Weight Initialization[mean = 0, std = 0.02]

# function to create discriminator model
def create_discriminator(image_size):
    weights_init = RandomNormal(mean=0, stddev=0.02)
    discriminator = Sequential([
        Input(shape=IMG_SIZE),
        SpectralNormalization(
            Conv2D(64, kernel_size=4, strides=2, padding='same',
                   kernel_initializer=weights_init)
        ),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        SpectralNormalization(
            Conv2D(128, kernel_size=4, strides=2, padding='same',
                   kernel_initializer=weights_init)
        ),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        SpectralNormalization(
            Conv2D(256, kernel_size=4, strides=2, padding='same',
                   kernel_initializer=weights_init)
        ),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        SpectralNormalization(
            Conv2D(512, kernel_size=4, strides=2, padding='same',
                   kernel_initializer=weights_init)
        ),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        GlobalMaxPooling2D(),
        Dense(1, activation='sigmoid'),
    ], name='discriminator_GAN')
    return discriminator


create_discriminator(image_size=IMG_SIZE).summary()

Model: "discriminator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 spectral_normalization_12 (  (None, 16, 16, 64)       3200      
 SpectralNormalization)                                          
                                                                 
 batch_normalization_154 (Ba  (None, 16, 16, 64)       256       
 tchNormalization)                                               
                                                                 
 leaky_re_lu_44 (LeakyReLU)  (None, 16, 16, 64)        0         
                                                                 
 spectral_normalization_13 (  (None, 8, 8, 128)        131328    
 SpectralNormalization)                                          
                                                                 
 batch_normalization_155 (Ba  (None, 8, 8, 128)        512       
 tchNormalization)                                               
                                                                 
 leaky_re_lu_45 (LeakyReLU)  (None, 8, 8, 128)         0         
                                                                 
 spectral_normalization_14 (  (None, 4, 4, 256)        524800    
 SpectralNormalization)                                          
                                                                 
 batch_normalization_156 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 leaky_re_lu_46 (LeakyReLU)  (None, 4, 4, 256)         0         
                                                                 
 spectral_normalization_15 (  (None, 2, 2, 512)        2098176   
 SpectralNormalization)                                          
                                                                 
 batch_normalization_157 (Ba  (None, 2, 2, 512)        2048      
 tchNormalization)                                               
                                                                 
 leaky_re_lu_47 (LeakyReLU)  (None, 2, 2, 512)         0         
                                                                 
 global_max_pooling2d (Globa  (None, 512)              0         
 lMaxPooling2D)                                                  
                                                                 
 dense_15 (Dense)            (None, 1)                 513       
                                                                 
=================================================================
Total params: 2,761,857
Trainable params: 2,758,977
Non-trainable params: 2,880
_________________________________________________________________

c:\Users\Soh Hong Yu\anaconda3\envs\gpu_env\lib\site-packages\keras\initializers\initializers_v2.py:120: UserWarning: The initializer RandomNormal is unseeded and being called multiple times, which will return identical values  each time (even if the initializer is unseeded). Please update your code to provide a seed to the initializer, or avoid using the same initalizer instance more than once.
  warnings.warn(

plot_model(create_discriminator(image_size=IMG_SIZE), to_file="images/models/DCGAN_discriminator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

DCGAN Generator

Unlike your typical neural architectures, the generator will have multiple convolutional upscaled.

1. Transposed Convolution Layers allows the normal convolution to go in the opposite direction. The main purpose is to allow the CNN to upsample from the noise latent vector to generated the images.
2. As sparse gradients will occur and diminish over time [random spikes in the loss during training], there will be many numerical operations performed on them. To fix this, we will add BatchNormalization to the network so that it can reduce covariate shifts.
3. He and Xavier weights initialization are unsuitable for training as it makes training more unstable due to the larger distribution of weights. To reduce the instability of training, we will be using Gaussian Weight Initialization[mean = 0, std = 0.02]
4. As we are using the LeakyReLU in the discriminator, and typically we will use ReLU in the generator. However, due to the sparse gradient during training and effects like the vanishing gradient problems, we can mitigate this effects using PReLU which as introduced with learning parameter to help training.

def create_generator(noise):

    # gaussian weights initialization
    weights_init = RandomNormal(mean=0, stddev=0.02)

    generator = Sequential([
        # fc block: handling latent vector z
        Input(shape=(noise,)),
        Dense(2*2*512),
        Reshape((2, 2, 512)),

        # Conv 1
        # using a kernel size that is a factor of the stride
        Conv2DTranspose(256, kernel_size=4, strides=2,
                        padding='same', kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        PReLU(),

        # Conv 2
        Conv2DTranspose(128, kernel_size=4, strides=2,
                        padding='same', kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        PReLU(),

        # Conv 3
        Conv2DTranspose(64, kernel_size=4, strides=2,
                        padding='same', kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        PReLU(),

        # Output Layer
        Conv2DTranspose(3, kernel_size=4, strides=2, padding='same',
                        activation='tanh', kernel_initializer=weights_init)
    ], name='generator_GAN')
    return generator


create_generator(noise=NOISE).summary()

Model: "generator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_16 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_6 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_18 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_16 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_6 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_18 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_158 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 p_re_lu_16 (PReLU)          (None, 4, 4, 256)         4096      
                                                                 
 conv2d_transpose_19 (Conv2D  (None, 8, 8, 128)        524416    
 Transpose)                                                      
                                                                 
 batch_normalization_159 (Ba  (None, 8, 8, 128)        512       
 tchNormalization)                                               
                                                                 
 p_re_lu_17 (PReLU)          (None, 8, 8, 128)         8192      
                                                                 
 conv2d_transpose_20 (Conv2D  (None, 16, 16, 64)       131136    
 Transpose)                                                      
                                                                 
 batch_normalization_160 (Ba  (None, 16, 16, 64)       256       
 tchNormalization)                                               
                                                                 
 p_re_lu_18 (PReLU)          (None, 16, 16, 64)        16384     
                                                                 
 conv2d_transpose_21 (Conv2D  (None, 32, 32, 3)        3075      
 Transpose)                                                      
                                                                 
=================================================================
Total params: 3,050,691
Trainable params: 3,049,795
Non-trainable params: 896
_________________________________________________________________

plot_model(create_generator(noise=NOISE), to_file="images/models/DCGAN_generator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

GAN Training Process

To train the GAN model, we will be making a class. As the training process is quite repetitive, we will be going to describe the process in a single iteration.

1. First the Generator will generate a batch of samples using a noise vector

2. Next the Discriminator will classify the real images and generated images [1 and 0].

3. Subsequently, the Discriminator will update the weights

   

4. The Generator will generate another batch of samples using the noise vector

5. Discriminator will compute a discrimination using the samples

6. The loss function is then applied between the discrimination computed previously

7. The Generator's weights will then be updated

To create the GAN Model, we will be referencing the TensorFlow website to help wrap a GAN model.

https://www.tensorflow.org/guide/keras/customizing_what_happens_in_fit#wrapping_up_an_end-to-end_gan_example

class GAN(Model):
    def __init__(self, discriminator, generator, noise):
        super(GAN, self).__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.noise = noise
        self.gen_loss_tracker = keras.metrics.Mean(name='generator_loss')
        self.disc_loss_tracker = keras.metrics.Mean(name='discriminator_loss')
        self.kl = tf.keras.metrics.KLDivergence()

    def compile(self, d_optimizer, g_optimizer, loss_fn):
        super(GAN, self).compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.loss_fn = loss_fn

    def train_step(self, real_images):
        if isinstance(real_images, tuple):
            real_images = real_images[0]
        # Sample random points in the latent space
        batch_size = tf.shape(real_images)[0]
        random_latent_vectors = tf.random.normal(
            shape=(batch_size, self.noise))
        # Decode them to fake images
        generated_images = self.generator(random_latent_vectors)
        # Combine them with real images
        combined_images = tf.concat([generated_images, real_images], axis=0)

        # Assemble labels discriminating real from fake images
        labels = tf.concat(
            [tf.zeros((batch_size, 1)), tf.ones((batch_size, 1))], axis=0
        )
        # Add random noise to the labels - important trick!
        labels += 0.05 * tf.random.uniform(tf.shape(labels))

        # Train the discriminator
        with tf.GradientTape() as tape:
            predictions = self.discriminator(combined_images)
            d_loss = self.loss_fn(labels, predictions)
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # TRAINING GENERATOR
        # Sample random points in the latent space
        random_latent_vectors = tf.random.normal(
            shape=(batch_size, self.noise))

        # Assemble labels that say "all real images"
        misleading_labels = tf.ones((batch_size, 1))

        # Train the generator (note that we should *not* update the weights
        # of the discriminator)!
        with tf.GradientTape() as tape:
            generated_images = self.generator(random_latent_vectors)
            predictions = self.discriminator(generated_images)
            g_loss = self.loss_fn(misleading_labels, predictions)
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(
            zip(grads, self.generator.trainable_weights))

        # Monitor Loss
        self.disc_loss_tracker.update_state(d_loss)
        self.gen_loss_tracker.update_state(g_loss)
        self.kl.update_state(y_true=real_images, y_pred=generated_images)

        return {
            "d_loss": self.disc_loss_tracker.result(),
            "g_loss": self.gen_loss_tracker.result(),
            "KL Divergence": self.kl.result(),
        }

DCGAN Callback

To help monitor the DCGAN's loss and progress and visualise the images after every patience. We will need to make a custom tensorflow callback that will help monitor the DCGAN.

class GANMonitor(keras.callbacks.Callback):
    def __init__(self, num_img=10, noise=128, patience=10, vmin=0, vmax=1):
        self.num_img = num_img
        self.noise = noise
        self.patience = patience
        self.vmin = vmin
        self.vmax = vmax
        self.constant_noise = tf.random.normal(
            shape=(self.num_img, self.noise))

    def generate_plot(self):
        # Generate Images
        generated_images = self.model.generator(self.constant_noise)
        # Normalise Image from [vmin, vmax] to [0, 1]
        generated_images -= self.vmin
        generated_images /= (self.vmax - self.vmin)
        row_size = int(np.ceil(self.num_img/5))
        fig = plt.figure(figsize=(10, 2*row_size), tight_layout=True)
        for i in range(self.num_img):
            ax = fig.add_subplot(row_size, 5, i+1)
            ax.imshow(generated_images[i])
            ax.axis('off')
        plt.show()

    def save_weights(self, epoch=None):
        try:
            if epoch != None:
                name = 'GAN/generator-epoch-{}.h5'.format(epoch)
                print('Generator Checkpoint - {}'.format(name))
                self.model.generator.save_weights(
                    filepath=name,
                    save_format='h5'
                )
        except Exception as e:
            print(e)

    def on_epoch_begin(self, epoch, logs=None):
        if epoch % self.patience == 0:
            self.generate_plot()
            self.save_weights(epoch)

    def on_train_end(self, epoch, logs=None):
        self.generate_plot()
        self.save_weights('Full Train')

callbacks = [
    GANMonitor(num_img=15, noise=128, patience=5, vmin=-1, vmax=1),
]

DCGAN Training

After setting the GAN architecture, we can now run the GAN models to generate images. First, we need to convert the data into a tensorflow dataset and training using multiprocessing to speed up process.

tf.keras.backend.clear_session()
K.clear_session()

gan = GAN(
    discriminator=create_discriminator(image_size=IMG_SIZE),
    generator=create_generator(noise=NOISE),
    noise=NOISE
)

gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=2e-4, beta_1=0.5),
    g_optimizer=keras.optimizers.Adam(learning_rate=2e-4, beta_1=0.5),
    loss_fn=keras.losses.BinaryCrossentropy(),
)

dataset = tf.data.Dataset.from_tensor_slices(x_train)
dataset = dataset.shuffle(buffer_size=BUFFER_SIZE).batch(
    batch_size=BATCH_SIZE, num_parallel_calls=AUTO).prefetch(AUTO)

gan_hist = gan.fit(dataset, epochs=EPOCHS, use_multiprocessing=True,
                   workers=16, callbacks=callbacks)

Generator Checkpoint - GAN/generator-epoch-0.h5
Epoch 1/100
782/782 [==============================] - 95s 116ms/step - d_loss: 0.8099 - g_loss: 2.3560 - KL Divergence: 5.5728
Epoch 2/100
782/782 [==============================] - 91s 116ms/step - d_loss: 0.5243 - g_loss: 1.3519 - KL Divergence: 5.0051
Epoch 3/100
782/782 [==============================] - 90s 115ms/step - d_loss: 0.5362 - g_loss: 1.3995 - KL Divergence: 4.9535
Epoch 4/100
782/782 [==============================] - 78s 99ms/step - d_loss: 0.4980 - g_loss: 1.6549 - KL Divergence: 5.0642
Epoch 5/100
782/782 [==============================] - 77s 98ms/step - d_loss: 0.4545 - g_loss: 1.7139 - KL Divergence: 4.9792

Generator Checkpoint - GAN/generator-epoch-5.h5
Epoch 6/100
782/782 [==============================] - 75s 96ms/step - d_loss: 0.4542 - g_loss: 1.7561 - KL Divergence: 5.2424
Epoch 7/100
782/782 [==============================] - 72s 92ms/step - d_loss: 0.4260 - g_loss: 1.8319 - KL Divergence: 5.3081
Epoch 8/100
782/782 [==============================] - 77s 99ms/step - d_loss: 0.3847 - g_loss: 1.9351 - KL Divergence: 5.3142
Epoch 9/100
782/782 [==============================] - 80s 103ms/step - d_loss: 0.4222 - g_loss: 1.7614 - KL Divergence: 5.2157
Epoch 10/100
782/782 [==============================] - 74s 94ms/step - d_loss: 0.4308 - g_loss: 1.6405 - KL Divergence: 5.2183

Generator Checkpoint - GAN/generator-epoch-10.h5
Epoch 11/100
782/782 [==============================] - 72s 93ms/step - d_loss: 0.4341 - g_loss: 1.5846 - KL Divergence: 5.2427
Epoch 12/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4353 - g_loss: 1.5529 - KL Divergence: 5.2889
Epoch 13/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4247 - g_loss: 1.6162 - KL Divergence: 5.2508
Epoch 14/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4312 - g_loss: 1.5784 - KL Divergence: 5.2568
Epoch 15/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.4375 - g_loss: 1.5366 - KL Divergence: 5.2752

Generator Checkpoint - GAN/generator-epoch-15.h5
Epoch 16/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4261 - g_loss: 1.5810 - KL Divergence: 5.3126
Epoch 17/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4012 - g_loss: 1.6312 - KL Divergence: 5.3076
Epoch 18/100
782/782 [==============================] - 66s 84ms/step - d_loss: 0.3830 - g_loss: 1.7145 - KL Divergence: 5.3971
Epoch 19/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.3504 - g_loss: 1.8088 - KL Divergence: 5.4369
Epoch 20/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.3303 - g_loss: 1.9310 - KL Divergence: 5.5025

Generator Checkpoint - GAN/generator-epoch-20.h5
Epoch 21/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.3051 - g_loss: 2.0590 - KL Divergence: 5.5511
Epoch 22/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2950 - g_loss: 2.1144 - KL Divergence: 5.5393
Epoch 23/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.3249 - g_loss: 2.0975 - KL Divergence: 5.4939
Epoch 24/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.2348 - g_loss: 2.2961 - KL Divergence: 5.5796
Epoch 25/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.2958 - g_loss: 2.2357 - KL Divergence: 5.5292

Generator Checkpoint - GAN/generator-epoch-25.h5
Epoch 26/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2135 - g_loss: 2.4048 - KL Divergence: 5.4618
Epoch 27/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.2243 - g_loss: 2.3866 - KL Divergence: 5.5467
Epoch 28/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.2290 - g_loss: 2.4840 - KL Divergence: 5.5828
Epoch 29/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.1924 - g_loss: 2.5711 - KL Divergence: 5.6314
Epoch 30/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2399 - g_loss: 2.5162 - KL Divergence: 5.5612

Generator Checkpoint - GAN/generator-epoch-30.h5
Epoch 31/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2327 - g_loss: 2.5737 - KL Divergence: 5.4968
Epoch 32/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1736 - g_loss: 2.7707 - KL Divergence: 5.5980
Epoch 33/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.1370 - g_loss: 2.8270 - KL Divergence: 5.5297
Epoch 34/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1860 - g_loss: 2.8384 - KL Divergence: 5.5760
Epoch 35/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2088 - g_loss: 2.8062 - KL Divergence: 5.4036

Generator Checkpoint - GAN/generator-epoch-35.h5
Epoch 36/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1440 - g_loss: 2.9751 - KL Divergence: 5.6307
Epoch 37/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1361 - g_loss: 2.9243 - KL Divergence: 5.4859
Epoch 38/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.0034 - g_loss: 2.9355 - KL Divergence: 5.6491
Epoch 39/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.1112 - g_loss: 3.0740 - KL Divergence: 5.4985
Epoch 40/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.1806 - g_loss: 3.0779 - KL Divergence: 5.4900

Generator Checkpoint - GAN/generator-epoch-40.h5
Epoch 41/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.2689 - g_loss: 2.9071 - KL Divergence: 5.3409
Epoch 42/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.6264 - g_loss: 2.7827 - KL Divergence: 5.4605
Epoch 43/100
782/782 [==============================] - 65s 84ms/step - d_loss: -0.7570 - g_loss: 2.6962 - KL Divergence: 5.4708
Epoch 44/100
782/782 [==============================] - 65s 84ms/step - d_loss: 0.0150 - g_loss: 2.8504 - KL Divergence: 5.3204
Epoch 45/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.3417 - g_loss: 2.6018 - KL Divergence: 5.2870

Generator Checkpoint - GAN/generator-epoch-45.h5
Epoch 46/100
782/782 [==============================] - 65s 83ms/step - d_loss: -1.5169 - g_loss: 2.7410 - KL Divergence: 5.4491
Epoch 47/100
782/782 [==============================] - 65s 84ms/step - d_loss: -1.1741 - g_loss: 2.4387 - KL Divergence: 5.5688
Epoch 48/100
782/782 [==============================] - 65s 83ms/step - d_loss: -0.7234 - g_loss: 2.8247 - KL Divergence: 5.3680
Epoch 49/100
782/782 [==============================] - 65s 84ms/step - d_loss: -3.7880 - g_loss: 3.0092 - KL Divergence: 5.3943
Epoch 50/100
782/782 [==============================] - 65s 83ms/step - d_loss: -1.7728 - g_loss: 2.7732 - KL Divergence: 5.2406

Generator Checkpoint - GAN/generator-epoch-50.h5
Epoch 51/100
782/782 [==============================] - 65s 83ms/step - d_loss: -1.4409 - g_loss: 2.9775 - KL Divergence: 5.2834
Epoch 52/100
782/782 [==============================] - 65s 83ms/step - d_loss: -7.8287 - g_loss: 3.3714 - KL Divergence: 5.3293
Epoch 53/100
782/782 [==============================] - 65s 83ms/step - d_loss: -7.6940 - g_loss: 3.7294 - KL Divergence: 5.2441
Epoch 54/100
782/782 [==============================] - 65s 83ms/step - d_loss: -5.6276 - g_loss: 3.3253 - KL Divergence: 5.1040
Epoch 55/100
782/782 [==============================] - 65s 83ms/step - d_loss: -6.5953 - g_loss: 4.3503 - KL Divergence: 5.0761

Generator Checkpoint - GAN/generator-epoch-55.h5
Epoch 56/100
782/782 [==============================] - 65s 83ms/step - d_loss: 3.8748 - g_loss: 4.5775 - KL Divergence: 5.1706
Epoch 57/100
782/782 [==============================] - 65s 83ms/step - d_loss: -1.5376 - g_loss: 3.7667 - KL Divergence: 5.2501
Epoch 58/100
782/782 [==============================] - 65s 83ms/step - d_loss: -9.7839 - g_loss: 4.6667 - KL Divergence: 5.1637
Epoch 59/100
782/782 [==============================] - 65s 83ms/step - d_loss: -17.7642 - g_loss: 6.0499 - KL Divergence: 5.1694
Epoch 60/100
782/782 [==============================] - 65s 82ms/step - d_loss: -1.4228 - g_loss: 6.8363 - KL Divergence: 5.0922

Generator Checkpoint - GAN/generator-epoch-60.h5
Epoch 61/100
782/782 [==============================] - 65s 83ms/step - d_loss: -21.4890 - g_loss: 5.6337 - KL Divergence: 5.1796
Epoch 62/100
782/782 [==============================] - 65s 83ms/step - d_loss: 0.4711 - g_loss: 6.6734 - KL Divergence: 5.1199
Epoch 63/100
782/782 [==============================] - 65s 84ms/step - d_loss: -2.3045 - g_loss: 6.2231 - KL Divergence: 5.0101
Epoch 64/100
782/782 [==============================] - 65s 83ms/step - d_loss: -2.6852 - g_loss: 6.2505 - KL Divergence: 4.9846
Epoch 65/100
782/782 [==============================] - 65s 83ms/step - d_loss: -8.7825 - g_loss: 7.3908 - KL Divergence: 5.1555

Generator Checkpoint - GAN/generator-epoch-65.h5
Epoch 66/100
782/782 [==============================] - 65s 83ms/step - d_loss: 1.6436 - g_loss: 7.2768 - KL Divergence: 5.1522
Epoch 67/100
782/782 [==============================] - 65s 83ms/step - d_loss: -21.2175 - g_loss: 7.4385 - KL Divergence: 5.1556
Epoch 68/100
782/782 [==============================] - 65s 83ms/step - d_loss: -56.5155 - g_loss: 9.7812 - KL Divergence: 5.1574
Epoch 69/100
782/782 [==============================] - 65s 83ms/step - d_loss: -24.4689 - g_loss: 9.5014 - KL Divergence: 5.1998
Epoch 70/100
782/782 [==============================] - 65s 83ms/step - d_loss: -68.9150 - g_loss: 11.8571 - KL Divergence: 5.0571

Generator Checkpoint - GAN/generator-epoch-70.h5
Epoch 71/100
782/782 [==============================] - 65s 83ms/step - d_loss: -40.4674 - g_loss: 12.1600 - KL Divergence: 5.0973
Epoch 72/100
782/782 [==============================] - 65s 83ms/step - d_loss: -54.0762 - g_loss: 15.8519 - KL Divergence: 5.0757
Epoch 73/100
782/782 [==============================] - 65s 83ms/step - d_loss: -102.1723 - g_loss: 13.6866 - KL Divergence: 5.0720
Epoch 74/100
782/782 [==============================] - 65s 83ms/step - d_loss: -3.3449 - g_loss: 20.1484 - KL Divergence: 5.1252
Epoch 75/100
782/782 [==============================] - 65s 83ms/step - d_loss: -30.8211 - g_loss: 16.9026 - KL Divergence: 5.0135

Generator Checkpoint - GAN/generator-epoch-75.h5
Epoch 76/100
782/782 [==============================] - 65s 83ms/step - d_loss: -134.2528 - g_loss: 18.6397 - KL Divergence: 5.0382
Epoch 77/100
782/782 [==============================] - 65s 83ms/step - d_loss: -25.5029 - g_loss: 23.1385 - KL Divergence: 5.0396
Epoch 78/100
782/782 [==============================] - 65s 83ms/step - d_loss: 8.7221 - g_loss: 22.2647 - KL Divergence: 5.3234
Epoch 79/100
782/782 [==============================] - 65s 83ms/step - d_loss: -27.8002 - g_loss: 21.4541 - KL Divergence: 5.1387
Epoch 80/100
782/782 [==============================] - 65s 83ms/step - d_loss: -39.1104 - g_loss: 22.9841 - KL Divergence: 5.0516

Generator Checkpoint - GAN/generator-epoch-80.h5
Epoch 81/100
782/782 [==============================] - 65s 84ms/step - d_loss: -65.5465 - g_loss: 26.5511 - KL Divergence: 5.1373
Epoch 82/100
782/782 [==============================] - 65s 84ms/step - d_loss: -45.9117 - g_loss: 24.1248 - KL Divergence: 5.1034
Epoch 83/100
782/782 [==============================] - 65s 84ms/step - d_loss: -145.6371 - g_loss: 26.2177 - KL Divergence: 5.1188
Epoch 84/100
782/782 [==============================] - 65s 83ms/step - d_loss: -40.8575 - g_loss: 31.3308 - KL Divergence: 5.0919
Epoch 85/100
782/782 [==============================] - 65s 83ms/step - d_loss: -150.8906 - g_loss: 29.5782 - KL Divergence: 5.0584

Generator Checkpoint - GAN/generator-epoch-85.h5
Epoch 86/100
782/782 [==============================] - 65s 83ms/step - d_loss: -208.3862 - g_loss: 42.6352 - KL Divergence: 5.2663
Epoch 87/100
782/782 [==============================] - 65s 83ms/step - d_loss: -123.9887 - g_loss: 37.4495 - KL Divergence: 5.0143
Epoch 88/100
782/782 [==============================] - 65s 83ms/step - d_loss: -38.3129 - g_loss: 46.1979 - KL Divergence: 4.9254
Epoch 89/100
782/782 [==============================] - 64s 82ms/step - d_loss: -30.6947 - g_loss: 37.9809 - KL Divergence: 5.1031
Epoch 90/100
782/782 [==============================] - 65s 83ms/step - d_loss: -133.7166 - g_loss: 36.0079 - KL Divergence: 5.0970

Generator Checkpoint - GAN/generator-epoch-90.h5
Epoch 91/100
782/782 [==============================] - 65s 83ms/step - d_loss: -164.8235 - g_loss: 42.4811 - KL Divergence: 5.0938
Epoch 92/100
782/782 [==============================] - 65s 83ms/step - d_loss: -173.5644 - g_loss: 46.2558 - KL Divergence: 5.0844
Epoch 93/100
782/782 [==============================] - 65s 83ms/step - d_loss: -243.3874 - g_loss: 52.1771 - KL Divergence: 5.0237
Epoch 94/100
782/782 [==============================] - 65s 83ms/step - d_loss: -8.8408 - g_loss: 60.3328 - KL Divergence: 5.0516
Epoch 95/100
782/782 [==============================] - 65s 83ms/step - d_loss: -203.5086 - g_loss: 60.0467 - KL Divergence: 5.1207

Generator Checkpoint - GAN/generator-epoch-95.h5
Epoch 96/100
782/782 [==============================] - 65s 83ms/step - d_loss: 80.9257 - g_loss: 81.5817 - KL Divergence: 4.8600
Epoch 97/100
782/782 [==============================] - 65s 83ms/step - d_loss: -47.3651 - g_loss: 53.4179 - KL Divergence: 4.9050
Epoch 98/100
782/782 [==============================] - 65s 83ms/step - d_loss: -51.3024 - g_loss: 51.2221 - KL Divergence: 4.9625
Epoch 99/100
782/782 [==============================] - 65s 83ms/step - d_loss: -260.3817 - g_loss: 56.3985 - KL Divergence: 5.0082
Epoch 100/100
782/782 [==============================] - 65s 83ms/step - d_loss: -75.4194 - g_loss: 59.0767 - KL Divergence: 5.1662

Generator Checkpoint - GAN/generator-epoch-Full Train.h5

Observations

We note that the images that generated at around 45 epochs are very clear with no clear signs of noise. But afterwards, we can see that noise starts showing up.

A few reasons why noise starts appearing in GAN includes:

1. Model collapse: When the generator produces a limited set of outputs, the model is unable to capture the diversity of the training data. This results in the generator producing samples that are variations of a few distinct modes or patterns, and the noise appearing in the generated images.

2. High variance in the data: GANs can struggle to generate high-quality images when there is a high variance in the training data. This can cause the generator to produce images with a lot of noise or artifacts.

3. Poorly chosen hyperparameters: GANs require careful tuning of hyperparameters such as learning rate, batch size, and network architecture. If the hyperparameters are not chosen correctly, the generator may produce images with a lot of noise.

4. Overfitting: GANs, like any other neural networks, can overfit to the training data, meaning that it will memorize the training examples and not generalize well to new examples. This can lead to the generator producing images with a lot of noise.

5. Unstable training: GANs can be difficult to train due to their instability during the training process. This can lead to the generator producing images with a lot of noise.

For this case, we can see that the model is issue is likely due to the model collapsing.

DCGAN Evaluation

As mentioned previously we will be using a quantitative and manual evaluation of the DCGAN.

We will start off with a quantitative analysis so that we can find the best model to generate the images for manual evaluation.

# story history object into dataframe
gan_hist_df = pd.DataFrame(gan_hist.history)

# using pandas dataframe to plot out learning curve
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 8), tight_layout=True)
gan_hist_df.loc[:, ["d_loss", 'g_loss']].plot(
    ax=ax1, title=r'Learning Curve of Loss Function CE')
gan_hist_df.loc[:, "KL Divergence"].plot(
    ax=ax2, title="Learning Curve of KL Divergence")
plt.show()

Observations

We notice that the model started collapsing after around 40 - 45 epochs as we can see the start of the divergence of g_loss and d_loss and the sudden drop in KL Divergence. From this, this shows that we will be using model epochs 40 to do a manual evaluation of the DCGAN model. We also note that by around 40 - 45 epochs the d_loss and g_loss starts to diverge from each other which suggest Model Collapsing occurred. This suggest that either the generator is too weak or the discriminator is too strong.

# Loading Weights for best Generator
saved_weights = 'GAN/generator-epoch-40.h5'
gan.generator.load_weights(saved_weights)
gan.generator.summary()

Model: "generator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_18 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_7 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_22 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_165 (Ba  (None, 4, 4, 256)        1024      
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_18 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_7 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_22 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_165 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 p_re_lu_19 (PReLU)          (None, 4, 4, 256)         4096      
                                                                 
 conv2d_transpose_23 (Conv2D  (None, 8, 8, 128)        524416    
 Transpose)                                                      
                                                                 
 batch_normalization_166 (Ba  (None, 8, 8, 128)        512       
 tchNormalization)                                               
                                                                 
 p_re_lu_20 (PReLU)          (None, 8, 8, 128)         8192      
                                                                 
 conv2d_transpose_24 (Conv2D  (None, 16, 16, 64)       131136    
 Transpose)                                                      
                                                                 
 batch_normalization_167 (Ba  (None, 16, 16, 64)       256       
 tchNormalization)                                               
                                                                 
 p_re_lu_21 (PReLU)          (None, 16, 16, 64)        16384     
                                                                 
 conv2d_transpose_25 (Conv2D  (None, 32, 32, 3)        3075      
 Transpose)                                                      
                                                                 
=================================================================
Total params: 3,050,691
Trainable params: 3,049,795
Non-trainable params: 896
_________________________________________________________________

# Generating 1000 Synthetic Images
random_noise = tf.random.normal(shape=(1000, NOISE))
synthetic_images = gan.generator.predict(random_noise)
print("Latent Vector Dim: {}\tGenerated Images Dim: {}".format(
    random_noise.shape, synthetic_images.shape))

# Scaling back to [0, 1]
synthetic_images -= -1
synthetic_images /= (1 - (-1))

# Display 15 randomly sampled images
fig = plt.figure(figsize=(10, 10), tight_layout=True)
for i in range(25):
    rand_idx = np.random.randint(0, len(synthetic_images))
    ax = fig.add_subplot(5, 5, i+1)
    ax.imshow(synthetic_images[rand_idx])
    ax.axis('off')
plt.show()

32/32 [==============================] - 4s 10ms/step
Latent Vector Dim: (1000, 128)	Generated Images Dim: (1000, 32, 32, 3)

Observations

We can see that the images are not clear and have a lot of different features of different class to help generate the images. It is hard to identify the type of images and which class it is trying to act like. Therefore models like Conditional GAN [cGAN] can help solve this issue.

gan_fid_class = GAN_FID(batch_size=512, noise=128,
                        sample_size=10000, buffer_size=1024)
gan_fid_class.fit(generator=gan.generator, train_data=x_train)
fid_score = gan_fid_class.evaluate()

Computing Real Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 9s 153ms/step
16/16 [==============================] - 3s 162ms/step
16/16 [==============================] - 3s 160ms/step
16/16 [==============================] - 3s 160ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 2s 154ms/step
16/16 [==============================] - 2s 156ms/step
16/16 [==============================] - 3s 160ms/step
16/16 [==============================] - 2s 136ms/step
16/16 [==============================] - 2s 139ms/step
16/16 [==============================] - 2s 139ms/step
16/16 [==============================] - 2s 139ms/step
16/16 [==============================] - 2s 140ms/step
16/16 [==============================] - 2s 137ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 141ms/step
16/16 [==============================] - 2s 140ms/step
16/16 [==============================] - 2s 138ms/step
16/16 [==============================] - 2s 141ms/step
Computing Generated Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 2s 141ms/step
16/16 [==============================] - 2s 140ms/step
16/16 [==============================] - 2s 141ms/step
16/16 [==============================] - 2s 137ms/step
16/16 [==============================] - 2s 141ms/step
16/16 [==============================] - 2s 143ms/step
16/16 [==============================] - 2s 142ms/step
16/16 [==============================] - 2s 140ms/step
16/16 [==============================] - 2s 142ms/step
16/16 [==============================] - 2s 145ms/step
16/16 [==============================] - 2s 141ms/step
16/16 [==============================] - 2s 143ms/step
16/16 [==============================] - 2s 143ms/step
16/16 [==============================] - 2s 153ms/step
16/16 [==============================] - 3s 168ms/step
16/16 [==============================] - 2s 156ms/step
16/16 [==============================] - 2s 153ms/step
16/16 [==============================] - 2s 149ms/step
16/16 [==============================] - 2s 149ms/step
16/16 [==============================] - 2s 151ms/step
Computed Embeddings	Real Images Embedding Shape: (10240, 2048)	Generated Images Embedding Shape: (10240, 2048)
The computed FID score is: 153.4580126583096

Observations

We see that the FID score is 153.4580126583096. This high FID score shows that the generated images are significantly different from the real images. Therefore more changes should be done to improve the FID score like using regularisation techniques or different architectures.

cGAN

cGAN also known as Conditional GAN allows the GAN model to control what type of images are generated. Therefore, in the cGAN model, labels will need to be parse into the model.

cGAN Discriminator

Basically, the discriminator is a binary classifier. For the cGAN model, we will make a typical CNN network. However, unlike a typical CNN network, we will be modifying it such that cGAN model will be improved.

1. Pooling Layers in a classic CNN are used to down sample activation features maps. However, this is not recommended for GAN as information will be loss. Therefore, we will be using a 2 x 2 Stride Convolution
2. Due to the vanishing gradient problem, the activation function for a GAN is usually a LeakyReLu. It allows gradients to flow easier through the architecture. This allows the function to compute the greatest value between the features and a small factor.
3. As sparse gradients will occur and diminish over time [random spikes in the loss during training], there will be many numerical operations performed on them. To fix this, we will add BatchNormalization to the network so that it can reduce covariate shifts. BatchNormalization will be used before non-Gaussian activations like LeakyReLU.

# function to create discriminator model
def create_cGAN_discriminator(image_size):
    # input image
    img_input = Input(shape=(image_size), name='Image_Input')
    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')
    base_discriminator = Sequential([
        Conv2D(64, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(128, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(256, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(512, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
    ], name='Base_Discriminator')
    discriminator = base_discriminator(img_input)

    discriminator = GlobalAveragePooling2D()(discriminator)

    # Concatenate - combine with conditions y
    merged_layer = Concatenate()([discriminator, conditions])
    discriminator = Dense(512, activation='relu')(merged_layer)

    # Output
    discriminator = Dense(1, activation='sigmoid',
                          name='Output_Layer')(discriminator)

    discriminator = Model(inputs=[img_input, conditions],
                          outputs=discriminator, name='discriminator_cGAN')
    return discriminator


create_cGAN_discriminator(image_size=IMG_SIZE).summary()

Model: "discriminator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
                                                                                                  
 global_average_pooling2d_7 (Gl  (None, 512)         0           ['Base_Discriminator[0][0]']     
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
                                                                                                  
 global_average_pooling2d_7 (Gl  (None, 512)         0           ['Base_Discriminator[0][0]']     
 obalAveragePooling2D)                                                                            
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_12 (Concatenate)   (None, 522)          0           ['global_average_pooling2d_7[0][0
                                                                 ]',                              
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_10 (Dense)               (None, 512)          267776      ['concatenate_12[0][0]']         
                                                                                                  
 Output_Layer (Dense)           (None, 1)            513         ['dense_10[0][0]']               
                                                                                                  
==================================================================================================
Total params: 3,028,673
Trainable params: 3,026,753
Non-trainable params: 1,920
__________________________________________________________________________________________________

plot_model(create_cGAN_discriminator(image_size=IMG_SIZE), to_file="images/models/cGAN_discriminator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

cGAN Generator

Unlike your typical neural architectures, the generator will have multiple convolutional upscaled.

1. Transposed Convolution Layers allows the normal convolution to go in the opposite direction. The main purpose is to allow the CNN to upsample from the noise latent vector to generated the images.
2. As sparse gradients will occur and diminish over time [random spikes in the loss during training], there will be many numerical operations performed on them. To fix this, we will add BatchNormalization to the network so that it can reduce covariate shifts.
3. Due to the sparse gradient during training and effects like the vanishing gradient problems, we can mitigate this effects using Leaky ReLU.

# function to create generator model
def create_cGAN_generator(noise):
    # latent noise vector z
    z = Input(shape=(noise,), name="Latent_Noise_Vector_z")

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    # Generator network
    merged_layer = Concatenate()([z, conditions])

    # FC: 2x2x512
    generator = Dense(2*2*512, activation='relu')(merged_layer)
    generator = BatchNormalization(momentum=0.9)(generator)
    generator = LeakyReLU(alpha=0.1)(generator)
    generator = Reshape((2, 2, 512))(generator)

    base_generator = Sequential([
        # Conv 1: 4x4x256
        Conv2DTranspose(256, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        # Conv 2: 8x8x128
        Conv2DTranspose(128, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        # Conv 3: 16x16x64
        Conv2DTranspose(64, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
    ], name='Base_Generator')
    generator = base_generator(generator)

    # Conv 4: 32x32x3
    generator = Conv2DTranspose(3, kernel_size=4, strides=2, padding='same',
                                activation='tanh', name='Output_Layer')(generator)

    generator = Model(inputs=[z, conditions],
                      outputs=generator,  name='generator_cGAN')
    return generator


create_cGAN_generator(noise=NOISE).summary()

Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_13 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_11 (Dense)               (None, 2048)         284672      ['concatenate_13[0][0]']         
                                                                                                  
 batch_normalization_138 (Batch  (None, 2048)        8192        ['dense_11[0][0]']               
 Normalization)                                                                                   
                                                                                                  
 leaky_re_lu_28 (LeakyReLU)     (None, 2048)         0           ['batch_normalization_138[0][0]']
                                                                                                  
 reshape_4 (Reshape)            (None, 2, 2, 512)    0           ['leaky_re_lu_28[0][0]']         
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2754752     ['reshape_4[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,050,691
Trainable params: 3,045,699
Non-trainable params: 4,992
__________________________________________________________________________________________________

plot_model(create_cGAN_generator(noise=NOISE), to_file="images/models/cGAN_generator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

cGAN Training Process

It is similar to the GAN Training process but with the inclusion of labels. To help us with the code, we will be referring to Keras example on a conditional GAN.

https://github.com/keras-team/keras-io/blob/master/examples/generative/conditional_gan.py

class ConditionalGAN(tf.keras.Model):
    def __init__(self, discriminator, generator, noise):
        super(ConditionalGAN, self).__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.noise = noise
        self.gen_loss_tracker = tf.keras.metrics.Mean(name="generator_loss")
        self.disc_loss_tracker = tf.keras.metrics.Mean(
            name="discriminator_loss")
        self.d_xy_tracker = tf.keras.metrics.Mean(name='Mean D(x|y)')
        self.d_g_zy_tracker = tf.keras.metrics.Mean(name='Mean D(G(z|y))')
        self.kl = tf.keras.metrics.KLDivergence()

    def compile(self, d_optimizer, g_optimizer, loss_fn):
        super(ConditionalGAN, self).compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.loss_fn = loss_fn

    def train_step(self, data):
        ### TRAINING DISCRIMINATOR ###
        # Unpack the data.
        real_images, condition = data

        # Sample for latent noise vector z
        batch_size = tf.shape(real_images)[0]
        latent_noise_vector = tf.random.normal(shape=(batch_size, self.noise))

        # Maps the noise latent vector and labels to generate fake images.
        generated_images = self.generator([latent_noise_vector, condition])

        # Combine with real images
        combined_images = tf.concat([generated_images, real_images], axis=0)
        combined_condition = tf.concat([condition, condition], axis=0)

        # Discrimination
        labels = tf.concat(
            [tf.zeros((batch_size, 1)), tf.ones((batch_size, 1))], axis=0
        )

        # Train the discriminator.
        with tf.GradientTape() as tape:
            first_predictions = self.discriminator(
                [combined_images, combined_condition])
            d_loss = self.loss_fn(labels, first_predictions)
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # Computing D(x|y)
        d_xy = tf.math.reduce_mean(first_predictions)

        ### TRAINING GENRATOR ###
        latent_noise_vector = tf.random.normal(shape=(batch_size, self.noise))

        # Assemble labels that say "all real images".
        misleading_labels = tf.ones((batch_size, 1))

        with tf.GradientTape() as tape:
            fake_images = self.generator([latent_noise_vector, condition])
            second_predictions = self.discriminator([fake_images, condition])
            g_loss = self.loss_fn(misleading_labels, second_predictions)
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(
            zip(grads, self.generator.trainable_weights))

        # Computing D(G(z|y))
        d_g_zy = tf.math.reduce_mean(second_predictions)

        # Monitor loss and metrics.
        self.gen_loss_tracker.update_state(g_loss)
        self.disc_loss_tracker.update_state(d_loss)
        self.d_xy_tracker.update_state(d_xy)
        self.d_g_zy_tracker.update_state(d_g_zy)
        self.kl.update_state(real_images, fake_images)

        return {
            "d_loss": self.disc_loss_tracker.result(),
            "g_loss": self.gen_loss_tracker.result(),
            "D(x|y)": self.d_xy_tracker.result(),
            "D(G(z|y))": self.d_g_zy_tracker.result(),
            "KL Divergence": self.
            kl.result(),
        }

cGAN Callback

To help monitor the cGAN's loss and progress and visualise the images after every patience. We will need to make a custom tensorflow callback that will help monitor the cGAN.

class cGANMonitor(keras.callbacks.Callback):
    def __init__(self, num_img=20, noise=128, patience=10, vmin=0, vmax=1):
        self.num_img = num_img
        self.noise = noise
        self.patience = patience
        self.vmin = vmin
        self.vmax = vmax
        self.latent_noise_vector = tf.random.normal(
            shape=(self.num_img, self.noise))
        self.conditions = to_categorical([0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                                          0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

    def generate_plot(self):
        # Generate Images
        generated_images = self.model.generator.predict(
            [self.latent_noise_vector, self.conditions])
        # Normalise Image from [vmin, vmax] to [0, 1]
        generated_images -= self.vmin
        generated_images /= (self.vmax - self.vmin)
        row_size = int(np.ceil(self.num_img/5))
        fig = plt.figure(figsize=(10, 2*row_size), tight_layout=True)
        for i in range(self.num_img):
            ax = fig.add_subplot(row_size, 5, i+1)
            ax.imshow(generated_images[i])
            ax.set_title(class_labels[i % 10])
            ax.axis('off')
        plt.show()

    def save_weights(self, epoch=None):
        try:
            if epoch != None:
                name = 'cGAN/generator-epoch-{}.h5'.format(epoch)
                print('Generator Checkpoint - {}'.format(name))
                self.model.generator.save_weights(
                    filepath=name,
                    save_format='h5'
                )
        except Exception as e:
            print(e)

    def on_epoch_begin(self, epoch, logs=None):
        if epoch % self.patience == 0:
            self.generate_plot()
            self.save_weights(epoch)

    def on_train_end(self, epoch, logs=None):
        self.generate_plot()
        self.save_weights('Full Train')

callbacks = [
    cGANMonitor(num_img=20, noise=128, patience=5, vmin=-1, vmax=1),
]

cGAN Training

After setting the cGAN architecture, we can now run the cGAN models to generate images. First, we need to convert the data into a tensorflow dataset and training using multiprocessing to speed up process.

tf.keras.backend.clear_session()
K.clear_session()

cond_gan = ConditionalGAN(
    discriminator=create_cGAN_discriminator(image_size=IMG_SIZE),
    generator=create_cGAN_generator(noise=NOISE),
    noise=NOISE
)

cond_gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    loss_fn=keras.losses.BinaryCrossentropy(),
)

dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
dataset = dataset.shuffle(buffer_size=1024).batch(
    BATCH_SIZE, num_parallel_calls=tf.data.AUTOTUNE).prefetch(tf.data.AUTOTUNE)

cond_gan_hist = cond_gan.fit(
    dataset, epochs=200, use_multiprocessing=True, workers=16, callbacks=callbacks)

1/1 [==============================] - 0s 355ms/step

Generator Checkpoint - cGAN/generator-epoch-0.h5
Epoch 1/200
782/782 [==============================] - 65s 79ms/step - d_loss: 0.3290 - g_loss: 2.8474 - D(x|y): 0.5407 - D(G(z|y)): 0.1412 - KL Divergence: 6.1464
Epoch 2/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.2402 - g_loss: 3.7414 - D(x|y): 0.4965 - D(G(z|y)): 0.0524 - KL Divergence: 5.9927
Epoch 3/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2934 - g_loss: 2.8933 - D(x|y): 0.4990 - D(G(z|y)): 0.0857 - KL Divergence: 4.8510
Epoch 4/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2417 - g_loss: 3.2650 - D(x|y): 0.5028 - D(G(z|y)): 0.0628 - KL Divergence: 4.4427
Epoch 5/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2386 - g_loss: 3.5025 - D(x|y): 0.5005 - D(G(z|y)): 0.0572 - KL Divergence: 4.5276
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-5.h5
Epoch 6/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2520 - g_loss: 3.2240 - D(x|y): 0.4999 - D(G(z|y)): 0.0781 - KL Divergence: 5.1005
Epoch 7/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2711 - g_loss: 3.0058 - D(x|y): 0.4993 - D(G(z|y)): 0.0978 - KL Divergence: 4.7024
Epoch 8/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2723 - g_loss: 2.9330 - D(x|y): 0.5011 - D(G(z|y)): 0.1089 - KL Divergence: 4.7046
Epoch 9/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2756 - g_loss: 2.8864 - D(x|y): 0.5011 - D(G(z|y)): 0.1193 - KL Divergence: 4.8275
Epoch 10/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2889 - g_loss: 2.6858 - D(x|y): 0.5009 - D(G(z|y)): 0.1370 - KL Divergence: 4.7402
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - cGAN/generator-epoch-10.h5
Epoch 11/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.3023 - g_loss: 2.6386 - D(x|y): 0.5003 - D(G(z|y)): 0.1455 - KL Divergence: 4.7275
Epoch 12/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.2889 - g_loss: 2.7427 - D(x|y): 0.4999 - D(G(z|y)): 0.1397 - KL Divergence: 4.7474
Epoch 13/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2824 - g_loss: 2.7091 - D(x|y): 0.5006 - D(G(z|y)): 0.1464 - KL Divergence: 4.7949
Epoch 14/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2738 - g_loss: 2.7683 - D(x|y): 0.5006 - D(G(z|y)): 0.1444 - KL Divergence: 4.7494
Epoch 15/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2598 - g_loss: 2.8938 - D(x|y): 0.5011 - D(G(z|y)): 0.1375 - KL Divergence: 4.7508
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-15.h5
Epoch 16/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2525 - g_loss: 2.9388 - D(x|y): 0.5001 - D(G(z|y)): 0.1374 - KL Divergence: 4.6120
Epoch 17/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2284 - g_loss: 3.1960 - D(x|y): 0.5008 - D(G(z|y)): 0.1217 - KL Divergence: 4.5727
Epoch 18/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.2133 - g_loss: 3.3671 - D(x|y): 0.5003 - D(G(z|y)): 0.1135 - KL Divergence: 4.4471
Epoch 19/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.2055 - g_loss: 3.4754 - D(x|y): 0.5003 - D(G(z|y)): 0.1114 - KL Divergence: 4.5915
Epoch 20/200
782/782 [==============================] - 61s 78ms/step - d_loss: 0.1963 - g_loss: 3.6419 - D(x|y): 0.5000 - D(G(z|y)): 0.1049 - KL Divergence: 4.5955
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - cGAN/generator-epoch-20.h5
Epoch 21/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1920 - g_loss: 3.7617 - D(x|y): 0.5006 - D(G(z|y)): 0.1020 - KL Divergence: 4.5135
Epoch 22/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1781 - g_loss: 3.8728 - D(x|y): 0.5001 - D(G(z|y)): 0.0979 - KL Divergence: 4.5717
Epoch 23/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1659 - g_loss: 4.1535 - D(x|y): 0.5003 - D(G(z|y)): 0.0881 - KL Divergence: 4.5492
Epoch 24/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1670 - g_loss: 4.1812 - D(x|y): 0.5001 - D(G(z|y)): 0.0890 - KL Divergence: 4.7242
Epoch 25/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.1561 - g_loss: 4.3943 - D(x|y): 0.5006 - D(G(z|y)): 0.0861 - KL Divergence: 4.7061
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - cGAN/generator-epoch-25.h5
Epoch 26/200
782/782 [==============================] - 72s 93ms/step - d_loss: 0.1571 - g_loss: 4.4908 - D(x|y): 0.5005 - D(G(z|y)): 0.0841 - KL Divergence: 4.6588
Epoch 27/200
782/782 [==============================] - 78s 100ms/step - d_loss: 0.1483 - g_loss: 4.6073 - D(x|y): 0.5001 - D(G(z|y)): 0.0808 - KL Divergence: 4.6768
Epoch 28/200
782/782 [==============================] - 84s 107ms/step - d_loss: 0.1423 - g_loss: 4.7443 - D(x|y): 0.5005 - D(G(z|y)): 0.0796 - KL Divergence: 4.6502
Epoch 29/200
782/782 [==============================] - 70s 90ms/step - d_loss: 0.1425 - g_loss: 4.8908 - D(x|y): 0.4995 - D(G(z|y)): 0.0765 - KL Divergence: 4.6786
Epoch 30/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.1297 - g_loss: 5.0276 - D(x|y): 0.5003 - D(G(z|y)): 0.0731 - KL Divergence: 4.6257
1/1 [==============================] - 0s 37ms/step

Generator Checkpoint - cGAN/generator-epoch-30.h5
Epoch 31/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1292 - g_loss: 5.0845 - D(x|y): 0.4998 - D(G(z|y)): 0.0723 - KL Divergence: 4.6513
Epoch 32/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.1969 - g_loss: 4.6288 - D(x|y): 0.5001 - D(G(z|y)): 0.0967 - KL Divergence: 4.7050
Epoch 33/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.1264 - g_loss: 5.3292 - D(x|y): 0.5008 - D(G(z|y)): 0.0679 - KL Divergence: 4.6659
Epoch 34/200
782/782 [==============================] - 79s 101ms/step - d_loss: 0.1197 - g_loss: 5.4663 - D(x|y): 0.4996 - D(G(z|y)): 0.0668 - KL Divergence: 4.6905
Epoch 35/200
782/782 [==============================] - 79s 101ms/step - d_loss: 0.1204 - g_loss: 5.5536 - D(x|y): 0.5002 - D(G(z|y)): 0.0642 - KL Divergence: 4.6716
1/1 [==============================] - 0s 31ms/step

Generator Checkpoint - cGAN/generator-epoch-35.h5
Epoch 36/200
782/782 [==============================] - 79s 101ms/step - d_loss: 0.1108 - g_loss: 5.5071 - D(x|y): 0.5002 - D(G(z|y)): 0.0636 - KL Divergence: 4.5958
Epoch 37/200
782/782 [==============================] - 78s 99ms/step - d_loss: 0.1098 - g_loss: 5.8188 - D(x|y): 0.5005 - D(G(z|y)): 0.0608 - KL Divergence: 4.7379
Epoch 38/200
782/782 [==============================] - 78s 99ms/step - d_loss: 0.1080 - g_loss: 5.9050 - D(x|y): 0.5006 - D(G(z|y)): 0.0562 - KL Divergence: 4.6329
Epoch 39/200
782/782 [==============================] - 79s 101ms/step - d_loss: 0.1068 - g_loss: 6.0401 - D(x|y): 0.5000 - D(G(z|y)): 0.0576 - KL Divergence: 4.4598
Epoch 40/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.1046 - g_loss: 5.9828 - D(x|y): 0.5007 - D(G(z|y)): 0.0591 - KL Divergence: 4.6524
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - cGAN/generator-epoch-40.h5
Epoch 41/200
782/782 [==============================] - 71s 90ms/step - d_loss: 0.1025 - g_loss: 6.0877 - D(x|y): 0.4998 - D(G(z|y)): 0.0571 - KL Divergence: 4.5610
Epoch 42/200
782/782 [==============================] - 74s 94ms/step - d_loss: 0.1025 - g_loss: 6.2132 - D(x|y): 0.5003 - D(G(z|y)): 0.0581 - KL Divergence: 4.6170
Epoch 43/200
782/782 [==============================] - 97s 124ms/step - d_loss: 0.1008 - g_loss: 6.3210 - D(x|y): 0.5001 - D(G(z|y)): 0.0544 - KL Divergence: 4.5913
Epoch 44/200
782/782 [==============================] - 125s 159ms/step - d_loss: 0.1061 - g_loss: 6.2544 - D(x|y): 0.5001 - D(G(z|y)): 0.0550 - KL Divergence: 4.5916
Epoch 45/200
782/782 [==============================] - 77s 99ms/step - d_loss: 0.0969 - g_loss: 6.4526 - D(x|y): 0.5004 - D(G(z|y)): 0.0523 - KL Divergence: 4.6210
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - cGAN/generator-epoch-45.h5
Epoch 46/200
782/782 [==============================] - 75s 96ms/step - d_loss: 0.0983 - g_loss: 6.5408 - D(x|y): 0.5004 - D(G(z|y)): 0.0500 - KL Divergence: 4.5318
Epoch 47/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0949 - g_loss: 6.6686 - D(x|y): 0.5002 - D(G(z|y)): 0.0504 - KL Divergence: 4.6332
Epoch 48/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.2141 - g_loss: 5.4477 - D(x|y): 0.5014 - D(G(z|y)): 0.0901 - KL Divergence: 4.7304
Epoch 49/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0843 - g_loss: 6.7116 - D(x|y): 0.5000 - D(G(z|y)): 0.0470 - KL Divergence: 4.7106
Epoch 50/200
782/782 [==============================] - 71s 90ms/step - d_loss: 0.0924 - g_loss: 6.7227 - D(x|y): 0.5001 - D(G(z|y)): 0.0501 - KL Divergence: 4.6189
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - cGAN/generator-epoch-50.h5
Epoch 51/200
782/782 [==============================] - 70s 90ms/step - d_loss: 0.0903 - g_loss: 6.8311 - D(x|y): 0.5002 - D(G(z|y)): 0.0485 - KL Divergence: 4.5783
Epoch 52/200
782/782 [==============================] - 73s 94ms/step - d_loss: 0.0870 - g_loss: 6.9528 - D(x|y): 0.5001 - D(G(z|y)): 0.0465 - KL Divergence: 4.4725
Epoch 53/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0865 - g_loss: 6.9041 - D(x|y): 0.4998 - D(G(z|y)): 0.0481 - KL Divergence: 4.4299
Epoch 54/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0868 - g_loss: 7.0610 - D(x|y): 0.4998 - D(G(z|y)): 0.0438 - KL Divergence: 4.4405
Epoch 55/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0894 - g_loss: 7.0356 - D(x|y): 0.4998 - D(G(z|y)): 0.0496 - KL Divergence: 4.5403
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - cGAN/generator-epoch-55.h5
Epoch 56/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0816 - g_loss: 7.2499 - D(x|y): 0.5003 - D(G(z|y)): 0.0415 - KL Divergence: 4.4725
Epoch 57/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0814 - g_loss: 7.3587 - D(x|y): 0.5000 - D(G(z|y)): 0.0422 - KL Divergence: 4.4042
Epoch 58/200
782/782 [==============================] - 73s 94ms/step - d_loss: 0.0778 - g_loss: 7.3986 - D(x|y): 0.5001 - D(G(z|y)): 0.0426 - KL Divergence: 4.5116
Epoch 59/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.0782 - g_loss: 7.2905 - D(x|y): 0.5002 - D(G(z|y)): 0.0435 - KL Divergence: 4.5159
Epoch 60/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0770 - g_loss: 7.4403 - D(x|y): 0.5003 - D(G(z|y)): 0.0422 - KL Divergence: 4.5372
1/1 [==============================] - 0s 40ms/step

Generator Checkpoint - cGAN/generator-epoch-60.h5
Epoch 61/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0815 - g_loss: 7.5033 - D(x|y): 0.4999 - D(G(z|y)): 0.0431 - KL Divergence: 4.3453
Epoch 62/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0812 - g_loss: 7.5420 - D(x|y): 0.5004 - D(G(z|y)): 0.0422 - KL Divergence: 4.4354
Epoch 63/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0775 - g_loss: 7.6043 - D(x|y): 0.5001 - D(G(z|y)): 0.0419 - KL Divergence: 4.4274
Epoch 64/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0742 - g_loss: 7.7003 - D(x|y): 0.4998 - D(G(z|y)): 0.0423 - KL Divergence: 4.4476
Epoch 65/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0726 - g_loss: 7.8266 - D(x|y): 0.5002 - D(G(z|y)): 0.0387 - KL Divergence: 4.3770
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - cGAN/generator-epoch-65.h5
Epoch 66/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0795 - g_loss: 7.7018 - D(x|y): 0.5000 - D(G(z|y)): 0.0380 - KL Divergence: 4.4405
Epoch 67/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.0758 - g_loss: 7.7602 - D(x|y): 0.5005 - D(G(z|y)): 0.0407 - KL Divergence: 4.5301
Epoch 68/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0706 - g_loss: 7.9645 - D(x|y): 0.5003 - D(G(z|y)): 0.0373 - KL Divergence: 4.4222
Epoch 69/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0718 - g_loss: 8.1038 - D(x|y): 0.5000 - D(G(z|y)): 0.0357 - KL Divergence: 4.4970
Epoch 70/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0751 - g_loss: 7.9405 - D(x|y): 0.5004 - D(G(z|y)): 0.0370 - KL Divergence: 4.5157
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-70.h5
Epoch 71/200
782/782 [==============================] - 59s 76ms/step - d_loss: 0.0748 - g_loss: 7.9451 - D(x|y): 0.4997 - D(G(z|y)): 0.0393 - KL Divergence: 4.5122
Epoch 72/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0688 - g_loss: 8.1707 - D(x|y): 0.5003 - D(G(z|y)): 0.0372 - KL Divergence: 4.5427
Epoch 73/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0668 - g_loss: 8.2270 - D(x|y): 0.5001 - D(G(z|y)): 0.0364 - KL Divergence: 4.4146
Epoch 74/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0685 - g_loss: 8.1333 - D(x|y): 0.4998 - D(G(z|y)): 0.0350 - KL Divergence: 4.5463
Epoch 75/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0692 - g_loss: 8.0679 - D(x|y): 0.4996 - D(G(z|y)): 0.0370 - KL Divergence: 4.5825
1/1 [==============================] - 0s 35ms/step

Generator Checkpoint - cGAN/generator-epoch-75.h5
Epoch 76/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0704 - g_loss: 8.0912 - D(x|y): 0.5001 - D(G(z|y)): 0.0369 - KL Divergence: 4.6126
Epoch 77/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0652 - g_loss: 8.3150 - D(x|y): 0.5000 - D(G(z|y)): 0.0353 - KL Divergence: 4.6889
Epoch 78/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0660 - g_loss: 8.3525 - D(x|y): 0.5005 - D(G(z|y)): 0.0345 - KL Divergence: 4.4762
Epoch 79/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0754 - g_loss: 8.2982 - D(x|y): 0.4997 - D(G(z|y)): 0.0380 - KL Divergence: 4.3649
Epoch 80/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0650 - g_loss: 8.3727 - D(x|y): 0.4999 - D(G(z|y)): 0.0355 - KL Divergence: 4.4189
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - cGAN/generator-epoch-80.h5
Epoch 81/200
782/782 [==============================] - 65s 84ms/step - d_loss: 0.0635 - g_loss: 8.5000 - D(x|y): 0.4999 - D(G(z|y)): 0.0334 - KL Divergence: 4.5453
Epoch 82/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0617 - g_loss: 8.5895 - D(x|y): 0.5002 - D(G(z|y)): 0.0337 - KL Divergence: 4.3807
Epoch 83/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0618 - g_loss: 8.5598 - D(x|y): 0.5004 - D(G(z|y)): 0.0326 - KL Divergence: 4.6094
Epoch 84/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0629 - g_loss: 8.6259 - D(x|y): 0.5001 - D(G(z|y)): 0.0332 - KL Divergence: 4.5424
Epoch 85/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0632 - g_loss: 8.6963 - D(x|y): 0.5003 - D(G(z|y)): 0.0314 - KL Divergence: 4.5816
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - cGAN/generator-epoch-85.h5
Epoch 86/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0584 - g_loss: 8.8067 - D(x|y): 0.5001 - D(G(z|y)): 0.0322 - KL Divergence: 4.4520
Epoch 87/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0642 - g_loss: 8.6613 - D(x|y): 0.5002 - D(G(z|y)): 0.0298 - KL Divergence: 4.4762
Epoch 88/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0601 - g_loss: 8.8067 - D(x|y): 0.5001 - D(G(z|y)): 0.0315 - KL Divergence: 4.4198
Epoch 89/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0585 - g_loss: 8.7492 - D(x|y): 0.5003 - D(G(z|y)): 0.0309 - KL Divergence: 4.5800
Epoch 90/200
782/782 [==============================] - 72s 93ms/step - d_loss: 0.0626 - g_loss: 8.8255 - D(x|y): 0.4996 - D(G(z|y)): 0.0306 - KL Divergence: 4.5677
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - cGAN/generator-epoch-90.h5
Epoch 91/200
782/782 [==============================] - 64s 81ms/step - d_loss: 0.0638 - g_loss: 8.7785 - D(x|y): 0.4999 - D(G(z|y)): 0.0345 - KL Divergence: 4.6347
Epoch 92/200
782/782 [==============================] - 64s 81ms/step - d_loss: 0.0586 - g_loss: 8.8176 - D(x|y): 0.5005 - D(G(z|y)): 0.0318 - KL Divergence: 4.5735
Epoch 93/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0558 - g_loss: 8.9601 - D(x|y): 0.5004 - D(G(z|y)): 0.0311 - KL Divergence: 4.5888
Epoch 94/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0605 - g_loss: 8.9461 - D(x|y): 0.5003 - D(G(z|y)): 0.0310 - KL Divergence: 4.5973
Epoch 95/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0575 - g_loss: 9.0080 - D(x|y): 0.5004 - D(G(z|y)): 0.0288 - KL Divergence: 4.5335
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - cGAN/generator-epoch-95.h5
Epoch 96/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0579 - g_loss: 8.8230 - D(x|y): 0.5002 - D(G(z|y)): 0.0306 - KL Divergence: 4.5771
Epoch 97/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0537 - g_loss: 9.0884 - D(x|y): 0.5002 - D(G(z|y)): 0.0297 - KL Divergence: 4.6725
Epoch 98/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0591 - g_loss: 9.0935 - D(x|y): 0.5001 - D(G(z|y)): 0.0303 - KL Divergence: 4.7145
Epoch 99/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0587 - g_loss: 9.0100 - D(x|y): 0.5000 - D(G(z|y)): 0.0277 - KL Divergence: 4.6664
Epoch 100/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0596 - g_loss: 9.0655 - D(x|y): 0.5000 - D(G(z|y)): 0.0321 - KL Divergence: 4.6086
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-100.h5
Epoch 101/200
782/782 [==============================] - 73s 94ms/step - d_loss: 0.0546 - g_loss: 9.2400 - D(x|y): 0.5001 - D(G(z|y)): 0.0277 - KL Divergence: 4.4846
Epoch 102/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0509 - g_loss: 9.3053 - D(x|y): 0.5002 - D(G(z|y)): 0.0281 - KL Divergence: 4.6203
Epoch 103/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.0579 - g_loss: 9.1309 - D(x|y): 0.5002 - D(G(z|y)): 0.0301 - KL Divergence: 4.5321
Epoch 104/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0558 - g_loss: 9.4042 - D(x|y): 0.5002 - D(G(z|y)): 0.0285 - KL Divergence: 4.5307
Epoch 105/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0576 - g_loss: 9.2568 - D(x|y): 0.5005 - D(G(z|y)): 0.0301 - KL Divergence: 4.6307
1/1 [==============================] - 0s 30ms/step

Generator Checkpoint - cGAN/generator-epoch-105.h5
Epoch 106/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0539 - g_loss: 9.3884 - D(x|y): 0.5001 - D(G(z|y)): 0.0293 - KL Divergence: 4.6269
Epoch 107/200
782/782 [==============================] - 70s 90ms/step - d_loss: 0.0525 - g_loss: 9.5311 - D(x|y): 0.4999 - D(G(z|y)): 0.0273 - KL Divergence: 4.5779
Epoch 108/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0596 - g_loss: 9.2875 - D(x|y): 0.5001 - D(G(z|y)): 0.0299 - KL Divergence: 4.5937
Epoch 109/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0524 - g_loss: 9.4086 - D(x|y): 0.5001 - D(G(z|y)): 0.0285 - KL Divergence: 4.6130
Epoch 110/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0496 - g_loss: 9.6614 - D(x|y): 0.5000 - D(G(z|y)): 0.0259 - KL Divergence: 4.5116
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - cGAN/generator-epoch-110.h5
Epoch 111/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0512 - g_loss: 9.4737 - D(x|y): 0.5003 - D(G(z|y)): 0.0263 - KL Divergence: 4.5772
Epoch 112/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0535 - g_loss: 9.5654 - D(x|y): 0.5004 - D(G(z|y)): 0.0265 - KL Divergence: 4.5738
Epoch 113/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0581 - g_loss: 9.4941 - D(x|y): 0.5002 - D(G(z|y)): 0.0285 - KL Divergence: 4.4917
Epoch 114/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0500 - g_loss: 9.7375 - D(x|y): 0.4999 - D(G(z|y)): 0.0255 - KL Divergence: 4.5181
Epoch 115/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.0523 - g_loss: 9.6659 - D(x|y): 0.5002 - D(G(z|y)): 0.0268 - KL Divergence: 4.4897
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - cGAN/generator-epoch-115.h5
Epoch 116/200
782/782 [==============================] - 69s 89ms/step - d_loss: 0.0515 - g_loss: 9.6950 - D(x|y): 0.4999 - D(G(z|y)): 0.0268 - KL Divergence: 4.4341
Epoch 117/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0538 - g_loss: 9.6219 - D(x|y): 0.5000 - D(G(z|y)): 0.0250 - KL Divergence: 4.3640
Epoch 118/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0542 - g_loss: 9.5436 - D(x|y): 0.5000 - D(G(z|y)): 0.0267 - KL Divergence: 4.4896
Epoch 119/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.0480 - g_loss: 9.7372 - D(x|y): 0.5002 - D(G(z|y)): 0.0265 - KL Divergence: 4.4639
Epoch 120/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0479 - g_loss: 10.0041 - D(x|y): 0.5002 - D(G(z|y)): 0.0249 - KL Divergence: 4.3933
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - cGAN/generator-epoch-120.h5
Epoch 121/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0490 - g_loss: 9.6872 - D(x|y): 0.5003 - D(G(z|y)): 0.0256 - KL Divergence: 4.4661
Epoch 122/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.0512 - g_loss: 9.5724 - D(x|y): 0.4997 - D(G(z|y)): 0.0272 - KL Divergence: 4.5358
Epoch 123/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0474 - g_loss: 9.8894 - D(x|y): 0.5002 - D(G(z|y)): 0.0237 - KL Divergence: 4.6033
Epoch 124/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0615 - g_loss: 9.4532 - D(x|y): 0.4997 - D(G(z|y)): 0.0309 - KL Divergence: 4.4342
Epoch 125/200
782/782 [==============================] - 73s 94ms/step - d_loss: 0.0454 - g_loss: 10.1005 - D(x|y): 0.5002 - D(G(z|y)): 0.0229 - KL Divergence: 4.4813
1/1 [==============================] - 0s 81ms/step

Generator Checkpoint - cGAN/generator-epoch-125.h5
Epoch 126/200
782/782 [==============================] - 75s 95ms/step - d_loss: 0.0433 - g_loss: 9.9961 - D(x|y): 0.5000 - D(G(z|y)): 0.0234 - KL Divergence: 4.6299
Epoch 127/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0496 - g_loss: 9.9761 - D(x|y): 0.5000 - D(G(z|y)): 0.0228 - KL Divergence: 4.4763
Epoch 128/200
782/782 [==============================] - 75s 96ms/step - d_loss: 0.0467 - g_loss: 9.9507 - D(x|y): 0.4999 - D(G(z|y)): 0.0252 - KL Divergence: 4.4770
Epoch 129/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0493 - g_loss: 9.9549 - D(x|y): 0.5003 - D(G(z|y)): 0.0244 - KL Divergence: 4.4190
Epoch 130/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0454 - g_loss: 10.0584 - D(x|y): 0.5002 - D(G(z|y)): 0.0228 - KL Divergence: 4.4198
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - cGAN/generator-epoch-130.h5
Epoch 131/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0494 - g_loss: 9.9569 - D(x|y): 0.5004 - D(G(z|y)): 0.0245 - KL Divergence: 4.3665
Epoch 132/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0478 - g_loss: 9.8345 - D(x|y): 0.5003 - D(G(z|y)): 0.0248 - KL Divergence: 4.4539
Epoch 133/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0487 - g_loss: 9.8149 - D(x|y): 0.5000 - D(G(z|y)): 0.0255 - KL Divergence: 4.5078
Epoch 134/200
782/782 [==============================] - 88s 112ms/step - d_loss: 0.0432 - g_loss: 10.2140 - D(x|y): 0.5000 - D(G(z|y)): 0.0228 - KL Divergence: 4.4664
Epoch 135/200
782/782 [==============================] - 109s 140ms/step - d_loss: 0.0471 - g_loss: 10.1809 - D(x|y): 0.5003 - D(G(z|y)): 0.0228 - KL Divergence: 4.5349
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - cGAN/generator-epoch-135.h5
Epoch 136/200
782/782 [==============================] - 91s 116ms/step - d_loss: 0.0536 - g_loss: 9.9969 - D(x|y): 0.5000 - D(G(z|y)): 0.0266 - KL Divergence: 4.5650
Epoch 137/200
782/782 [==============================] - 97s 124ms/step - d_loss: 0.0422 - g_loss: 10.0192 - D(x|y): 0.5001 - D(G(z|y)): 0.0221 - KL Divergence: 4.5449
Epoch 138/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0488 - g_loss: 9.9873 - D(x|y): 0.5001 - D(G(z|y)): 0.0231 - KL Divergence: 4.4575
Epoch 139/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0406 - g_loss: 10.1816 - D(x|y): 0.5001 - D(G(z|y)): 0.0207 - KL Divergence: 4.4748
Epoch 140/200
782/782 [==============================] - 93s 119ms/step - d_loss: 0.0412 - g_loss: 10.1282 - D(x|y): 0.5002 - D(G(z|y)): 0.0245 - KL Divergence: 4.3233
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - cGAN/generator-epoch-140.h5
Epoch 141/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0440 - g_loss: 10.2198 - D(x|y): 0.5002 - D(G(z|y)): 0.0218 - KL Divergence: 4.5105
Epoch 142/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0449 - g_loss: 10.2572 - D(x|y): 0.5001 - D(G(z|y)): 0.0215 - KL Divergence: 4.4829
Epoch 143/200
782/782 [==============================] - 45s 57ms/step - d_loss: 0.0396 - g_loss: 10.2141 - D(x|y): 0.5000 - D(G(z|y)): 0.0222 - KL Divergence: 4.3987
Epoch 144/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0434 - g_loss: 10.3029 - D(x|y): 0.5004 - D(G(z|y)): 0.0222 - KL Divergence: 4.5692
Epoch 145/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0450 - g_loss: 10.2961 - D(x|y): 0.4997 - D(G(z|y)): 0.0230 - KL Divergence: 4.3560
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - cGAN/generator-epoch-145.h5
Epoch 146/200
782/782 [==============================] - 50s 64ms/step - d_loss: 0.0439 - g_loss: 10.3285 - D(x|y): 0.5004 - D(G(z|y)): 0.0213 - KL Divergence: 4.3877
Epoch 147/200
782/782 [==============================] - 50s 64ms/step - d_loss: 0.0399 - g_loss: 10.2543 - D(x|y): 0.4999 - D(G(z|y)): 0.0219 - KL Divergence: 4.5931
Epoch 148/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0463 - g_loss: 10.0871 - D(x|y): 0.5002 - D(G(z|y)): 0.0223 - KL Divergence: 4.5000
Epoch 149/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0450 - g_loss: 10.0692 - D(x|y): 0.5003 - D(G(z|y)): 0.0210 - KL Divergence: 4.5824
Epoch 150/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0429 - g_loss: 10.1244 - D(x|y): 0.4999 - D(G(z|y)): 0.0209 - KL Divergence: 4.4747
1/1 [==============================] - 0s 31ms/step

Generator Checkpoint - cGAN/generator-epoch-150.h5
Epoch 151/200
782/782 [==============================] - 50s 64ms/step - d_loss: 0.0424 - g_loss: 10.0620 - D(x|y): 0.5004 - D(G(z|y)): 0.0207 - KL Divergence: 5.0981
Epoch 152/200
782/782 [==============================] - 53s 68ms/step - d_loss: 0.0480 - g_loss: 9.8623 - D(x|y): 0.5000 - D(G(z|y)): 0.0220 - KL Divergence: 4.7471
Epoch 153/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0478 - g_loss: 9.7459 - D(x|y): 0.5004 - D(G(z|y)): 0.0228 - KL Divergence: 4.8388
Epoch 154/200
782/782 [==============================] - 80s 103ms/step - d_loss: 0.0473 - g_loss: 9.6613 - D(x|y): 0.5002 - D(G(z|y)): 0.0213 - KL Divergence: 4.7579
Epoch 155/200
782/782 [==============================] - 79s 101ms/step - d_loss: 0.0494 - g_loss: 9.6006 - D(x|y): 0.5005 - D(G(z|y)): 0.0223 - KL Divergence: 4.4742
1/1 [==============================] - 0s 36ms/step

Generator Checkpoint - cGAN/generator-epoch-155.h5
Epoch 156/200
782/782 [==============================] - 75s 96ms/step - d_loss: 0.0507 - g_loss: 9.4706 - D(x|y): 0.5001 - D(G(z|y)): 0.0226 - KL Divergence: 4.4549
Epoch 157/200
782/782 [==============================] - 78s 99ms/step - d_loss: 0.0516 - g_loss: 9.5134 - D(x|y): 0.5002 - D(G(z|y)): 0.0234 - KL Divergence: 4.9313
Epoch 158/200
782/782 [==============================] - 89s 114ms/step - d_loss: 0.0515 - g_loss: 9.5337 - D(x|y): 0.5005 - D(G(z|y)): 0.0237 - KL Divergence: 4.3586
Epoch 159/200
782/782 [==============================] - 93s 119ms/step - d_loss: 0.0576 - g_loss: 9.4633 - D(x|y): 0.5001 - D(G(z|y)): 0.0243 - KL Divergence: 4.4471
Epoch 160/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0534 - g_loss: 9.3525 - D(x|y): 0.5007 - D(G(z|y)): 0.0256 - KL Divergence: 4.2428
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - cGAN/generator-epoch-160.h5
Epoch 161/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0553 - g_loss: 9.4093 - D(x|y): 0.5002 - D(G(z|y)): 0.0246 - KL Divergence: 4.6710
Epoch 162/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0495 - g_loss: 9.4292 - D(x|y): 0.5003 - D(G(z|y)): 0.0230 - KL Divergence: 4.6330
Epoch 163/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0573 - g_loss: 9.5676 - D(x|y): 0.5003 - D(G(z|y)): 0.0242 - KL Divergence: 4.6990
Epoch 164/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0536 - g_loss: 9.3422 - D(x|y): 0.5005 - D(G(z|y)): 0.0239 - KL Divergence: 4.7132
Epoch 165/200
782/782 [==============================] - 61s 78ms/step - d_loss: 0.0580 - g_loss: 9.3585 - D(x|y): 0.4996 - D(G(z|y)): 0.0257 - KL Divergence: 4.4489
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - cGAN/generator-epoch-165.h5
Epoch 166/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0587 - g_loss: 9.3344 - D(x|y): 0.5001 - D(G(z|y)): 0.0234 - KL Divergence: 4.6203
Epoch 167/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.0522 - g_loss: 9.3847 - D(x|y): 0.5005 - D(G(z|y)): 0.0229 - KL Divergence: 4.2342
Epoch 168/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.0553 - g_loss: 9.4722 - D(x|y): 0.5001 - D(G(z|y)): 0.0244 - KL Divergence: 4.6620
Epoch 169/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0536 - g_loss: 9.3885 - D(x|y): 0.5002 - D(G(z|y)): 0.0240 - KL Divergence: 4.3944
Epoch 170/200
782/782 [==============================] - 62s 80ms/step - d_loss: 0.0527 - g_loss: 9.3972 - D(x|y): 0.5002 - D(G(z|y)): 0.0232 - KL Divergence: 4.8576
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - cGAN/generator-epoch-170.h5
Epoch 171/200
782/782 [==============================] - 62s 80ms/step - d_loss: 0.0585 - g_loss: 9.3405 - D(x|y): 0.5002 - D(G(z|y)): 0.0260 - KL Divergence: 5.3460
Epoch 172/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.0561 - g_loss: 9.4856 - D(x|y): 0.5004 - D(G(z|y)): 0.0253 - KL Divergence: 4.6888
Epoch 173/200
782/782 [==============================] - 62s 79ms/step - d_loss: 0.0562 - g_loss: 9.3107 - D(x|y): 0.5003 - D(G(z|y)): 0.0241 - KL Divergence: 4.5992
Epoch 174/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.0513 - g_loss: 9.4123 - D(x|y): 0.5000 - D(G(z|y)): 0.0230 - KL Divergence: 4.7673
Epoch 175/200
782/782 [==============================] - 63s 80ms/step - d_loss: 0.0531 - g_loss: 9.2189 - D(x|y): 0.5004 - D(G(z|y)): 0.0233 - KL Divergence: 4.8245
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-175.h5
Epoch 176/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0624 - g_loss: 9.4788 - D(x|y): 0.5005 - D(G(z|y)): 0.0268 - KL Divergence: 4.2699
Epoch 177/200
782/782 [==============================] - 63s 81ms/step - d_loss: 0.0489 - g_loss: 9.4795 - D(x|y): 0.5000 - D(G(z|y)): 0.0224 - KL Divergence: 4.0231
Epoch 178/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0541 - g_loss: 9.6320 - D(x|y): 0.5002 - D(G(z|y)): 0.0226 - KL Divergence: 4.0148
Epoch 179/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.0585 - g_loss: 9.5654 - D(x|y): 0.5005 - D(G(z|y)): 0.0239 - KL Divergence: 4.5631
Epoch 180/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.0527 - g_loss: 9.3711 - D(x|y): 0.4999 - D(G(z|y)): 0.0226 - KL Divergence: 4.5931
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - cGAN/generator-epoch-180.h5
Epoch 181/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.0566 - g_loss: 9.3943 - D(x|y): 0.5002 - D(G(z|y)): 0.0237 - KL Divergence: 4.4692
Epoch 182/200
782/782 [==============================] - 74s 94ms/step - d_loss: 0.0566 - g_loss: 9.4189 - D(x|y): 0.5002 - D(G(z|y)): 0.0227 - KL Divergence: 4.7486
Epoch 183/200
782/782 [==============================] - 77s 98ms/step - d_loss: 0.0587 - g_loss: 9.3143 - D(x|y): 0.5001 - D(G(z|y)): 0.0255 - KL Divergence: 4.5081
Epoch 184/200
782/782 [==============================] - 70s 90ms/step - d_loss: 0.0562 - g_loss: 9.3613 - D(x|y): 0.5003 - D(G(z|y)): 0.0254 - KL Divergence: 4.7135
Epoch 185/200
782/782 [==============================] - 76s 97ms/step - d_loss: 0.0528 - g_loss: 9.6380 - D(x|y): 0.5004 - D(G(z|y)): 0.0232 - KL Divergence: 4.3025
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - cGAN/generator-epoch-185.h5
Epoch 186/200
782/782 [==============================] - 79s 102ms/step - d_loss: 0.0624 - g_loss: 8.7457 - D(x|y): 0.4998 - D(G(z|y)): 0.0273 - KL Divergence: 4.3775
Epoch 187/200
782/782 [==============================] - 88s 113ms/step - d_loss: 0.0487 - g_loss: 9.3505 - D(x|y): 0.5001 - D(G(z|y)): 0.0234 - KL Divergence: 4.2476
Epoch 188/200
782/782 [==============================] - 93s 119ms/step - d_loss: 0.0526 - g_loss: 9.3762 - D(x|y): 0.5002 - D(G(z|y)): 0.0234 - KL Divergence: 4.6776
Epoch 189/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.0529 - g_loss: 9.6369 - D(x|y): 0.5002 - D(G(z|y)): 0.0231 - KL Divergence: 4.9668
Epoch 190/200
782/782 [==============================] - 88s 113ms/step - d_loss: 0.0513 - g_loss: 9.3026 - D(x|y): 0.5001 - D(G(z|y)): 0.0224 - KL Divergence: 4.3542
1/1 [==============================] - 0s 33ms/step

Generator Checkpoint - cGAN/generator-epoch-190.h5
Epoch 191/200
782/782 [==============================] - 86s 109ms/step - d_loss: 0.0515 - g_loss: 9.3558 - D(x|y): 0.5001 - D(G(z|y)): 0.0251 - KL Divergence: 4.4981
Epoch 192/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.0554 - g_loss: 9.4918 - D(x|y): 0.5001 - D(G(z|y)): 0.0230 - KL Divergence: 4.1773
Epoch 193/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.0611 - g_loss: 9.2813 - D(x|y): 0.5002 - D(G(z|y)): 0.0246 - KL Divergence: 4.9039
Epoch 194/200
782/782 [==============================] - 69s 89ms/step - d_loss: 0.0470 - g_loss: 8.6261 - D(x|y): 0.4999 - D(G(z|y)): 0.0234 - KL Divergence: 4.4972
Epoch 195/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.0519 - g_loss: 9.2205 - D(x|y): 0.5009 - D(G(z|y)): 0.0226 - KL Divergence: 4.6163
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - cGAN/generator-epoch-195.h5
Epoch 196/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0538 - g_loss: 9.2810 - D(x|y): 0.5002 - D(G(z|y)): 0.0236 - KL Divergence: 4.8282
Epoch 197/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0523 - g_loss: 9.2768 - D(x|y): 0.5004 - D(G(z|y)): 0.0235 - KL Divergence: 4.2958
Epoch 198/200
782/782 [==============================] - 76s 97ms/step - d_loss: 0.0522 - g_loss: 9.3172 - D(x|y): 0.5006 - D(G(z|y)): 0.0218 - KL Divergence: 4.7974
Epoch 199/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.0593 - g_loss: 9.5241 - D(x|y): 0.5004 - D(G(z|y)): 0.0246 - KL Divergence: 4.5065
Epoch 200/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.0547 - g_loss: 9.3898 - D(x|y): 0.5004 - D(G(z|y)): 0.0241 - KL Divergence: 4.4251
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - cGAN/generator-epoch-Full Train.h5

Observations

We notice that around 140 - 145 epochs, the automobile class starts to show the same images despite having the different noises. The model has learned a particular distribution of data which will generate a monotonous output [same images]. After more training certain labels like cats and frog look better and generate better results.

cGAN Evaluation

As mentioned previously we will be using a quantitative and manual evaluation of the cGAN.

We will start off with a quantitative analysis so that we can find the best model to generate the images for manual evaluation.

# story history object into dataframe
cond_gan_hist_df = pd.DataFrame(cond_gan_hist.history)

# using pandas dataframe to plot out learning curve
fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(24, 8), tight_layout=True)
cond_gan_hist_df.loc[:, ["d_loss", 'g_loss']].plot(
    ax=ax1, title=r'Learning Curve of Loss Function CE')
cond_gan_hist_df.loc[:, ['D(x|y)', 'D(G(z|y))']].plot(
    ax=ax2, title=r'Learning Curve of Discriminator')
cond_gan_hist_df.loc[:, 'KL Divergence'].plot(
    ax=ax3, title=r'Learning Curve of Probabilistic Similarity')
plt.show()

Observations

We notice that the around 140 - 145 epochs, the model starts collapsing where the images starts repeating. This is seen from the learning curve where the g_loss started dropping. We also note that after 145 epochs the KL divergences starts fluctuating rapidly. To choose best model, we will choose the lowest KL divergence before 140 epochs which is 130 epochs.

# Loading Weights for best Generator
saved_weights = 'cGAN/generator-epoch-130.h5'
cond_gan.generator.load_weights(saved_weights)
cond_gan.generator.summary()

Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_5 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_7 (Dense)                (None, 2048)         284672      ['concatenate_5[0][0]']          
                                                                                                  
 batch_normalization_120 (Batch  (None, 2048)        8192        ['dense_7[0][0]']                
 Normalization)                                                                                   
                                                                                                  
 leaky_re_lu_20 (LeakyReLU)     (None, 2048)         0           ['batch_normalization_120[0][0]']
                                                                                                  
 reshape_3 (Reshape)            (None, 2, 2, 512)    0           ['leaky_re_lu_20[0][0]']         
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2754752     ['reshape_3[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,050,691
Trainable params: 3,045,699
Non-trainable params: 4,992
__________________________________________________________________________________________________

n = 10000

# generating labels
labels = np.random.randint(low=0, high=10, size=n)
one_hot_labels = to_categorical(labels)

# Generating 1000 Synthetic Images
random_noise = tf.random.normal(shape=(n, NOISE))

synthetic_images = cond_gan.generator.predict(
    [random_noise, one_hot_labels])
print("Latent Vector Dim: {}\tGenerated Images Dim: {}".format(
    random_noise.shape, synthetic_images.shape))

# Scaling back to [0, 1]
synthetic_images -= -1
synthetic_images /= (1 - (-1))

# Display 25 randomly sampled images
fig = plt.figure(figsize=(10, 10), tight_layout=True)
for i in range(25):
    rand_idx = np.random.randint(0, len(synthetic_images))
    ax = fig.add_subplot(5, 5, i+1)
    ax.imshow(synthetic_images[rand_idx])
    ax.set_title(class_labels[labels[rand_idx]])
    ax.axis('off')
plt.show()

313/313 [==============================] - 3s 7ms/step
Latent Vector Dim: (10000, 128)	Generated Images Dim: (10000, 32, 32, 3)

Observations

We can see that the model is able to clearly generate vehicles with good precision but the animals are more jumbled up. Features like the texture can be seen in some of the images which suggest the model should want to have more training to help make the model better. But due to model collapsing, changes needed to be made to the architecture and using dropout and regularisation methods can make the model better.

cond_gan_fid_class = GAN_FID(batch_size=512, noise=128,
                             sample_size=10000, buffer_size=1024, labels=True)
cond_gan_fid_class.fit(generator=cond_gan.generator, train_data=x_train)
fid_score = cond_gan_fid_class.evaluate()

Computing Real Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 4s 133ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 134ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 134ms/step
16/16 [==============================] - 2s 133ms/step
Computing Generated Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 153ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 144ms/step
16/16 [==============================] - 2s 155ms/step
16/16 [==============================] - 3s 161ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 162ms/step
16/16 [==============================] - 3s 168ms/step
16/16 [==============================] - 3s 160ms/step
16/16 [==============================] - 3s 158ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 164ms/step
Computed Embeddings	Real Images Embedding Shape: (10240, 2048)	Generated Images Embedding Shape: (10240, 2048)
The computed FID score is: 94.536161025068

Observations

We see that the FID score has significantly reduced from 153.4580126583096 to 94.536161025068 [DCGAN to cGAN]. Although the FID score is still high but the score is below 100, we can see that with the addition of labels the model is more closely resembling the training data but there are still some noticeable differences.

ACGAN

Auxiliary Classifier Generative Adversarial Network, or ACGAN in short, follows a similar architecture to a Conditional GAN. There are two main difference between the two architectures - the discriminator produces two outputs and an auxiliary loss function is included in the training process.

Therefore, we will be only modifying the discriminator.

ACGAN Discriminator

As the ACGAN is very similar to the cGAN, we will be only modifying the last output layer. The ACGAN wants to not only lower the loss between real and fake images, but also reduce the loss between the conditions and the predicted conditions.

# function to create discriminator model
def create_ACGAN_discriminator(image_size):
    # input image
    img_input = Input(shape=(image_size), name='Image_Input')

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    base_discriminator = Sequential([
        Conv2D(64, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(128, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(256, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
        Conv2D(512, kernel_size=4, strides=2, padding='same'),
        BatchNormalization(momentum=0.9),
        LeakyReLU(alpha=0.1),
    ], name='Base_Discriminator')
    discriminator = base_discriminator(img_input)

    discriminator = GlobalAveragePooling2D()(discriminator)

    # Concatenate - combine with conditions y
    merged_layer = Concatenate()([discriminator, conditions])
    discriminator = Dense(512, activation='relu')(merged_layer)

    # Output
    Disc_Output_Layer = Dense(1, activation='sigmoid',
                              name='Disc_Output_Layer')(discriminator)
    Aux_Output_Layer = Dense(10, activation='softmax',
                             name='Aux_Output_Layer')(discriminator)

    discriminator = Model(inputs=[img_input, conditions],
                          outputs=[Disc_Output_Layer, Aux_Output_Layer], name='discriminator_ACGAN')
    return discriminator


create_ACGAN_discriminator(image_size=IMG_SIZE).summary()

Model: "discriminator_ACGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
                                                                                                  
 global_average_pooling2d_8 (Gl  (None, 512)         0           ['Base_Discriminator[0][0]']     
 obalAveragePooling2D)                                                                            
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_14 (Concatenate)   (None, 522)          0           ['global_average_pooling2d_8[0][0
                                                                 ]',                              
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_12 (Dense)               (None, 512)          267776      ['concatenate_14[0][0]']         
                                                                                                  
 Disc_Output_Layer (Dense)      (None, 1)            513         ['dense_12[0][0]']               
                                                                                                  
 Aux_Output_Layer (Dense)       (None, 10)           5130        ['dense_12[0][0]']               
                                                                                                  
==================================================================================================
Total params: 3,033,803
Trainable params: 3,031,883
Non-trainable params: 1,920
__________________________________________________________________________________________________

plot_model(create_ACGAN_discriminator(image_size=IMG_SIZE), to_file="images/models/ACGAN_discriminator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

ACGAN Training Process

It is similar to the cGAN Training process but with a new loss function. To help us with the code, we will be referring to Keras example on a conditional GAN.

https://github.com/keras-team/keras-io/blob/master/examples/generative/conditional_gan.py

class AConditionalGAN(tf.keras.Model):
    def __init__(self, discriminator, generator, noise):
        super(AConditionalGAN, self).__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.noise = noise
        self.gen_loss_tracker = tf.keras.metrics.Mean(name="generator_loss")
        self.disc_loss_tracker = tf.keras.metrics.Mean(
            name="discriminator_loss")
        self.d_xy_tracker = tf.keras.metrics.Mean(name='Mean D(x|y)')
        self.d_g_zy_tracker = tf.keras.metrics.Mean(name='Mean D(G(z|y))')
        self.kl = tf.keras.metrics.KLDivergence()

    def compile(self, d_optimizer, g_optimizer, disc_loss_fn, aux_loss_fn):
        super(AConditionalGAN, self).compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.disc_loss_fn = disc_loss_fn
        self.aux_loss_fn = aux_loss_fn

    def train_step(self, data):
        ### TRAINING DISCRIMINATOR ###
        # Unpack the data.
        real_images, condition = data

        # Sample for latent noise vector z
        batch_size = tf.shape(real_images)[0]
        latent_noise_vector = tf.random.normal(
            shape=(batch_size, self.noise))

        # Maps the noise latent vector and labels to generate fake images.
        generated_images = self.generator([latent_noise_vector, condition])

        # Combine with real images
        combined_images = tf.concat([generated_images, real_images], axis=0)
        combined_condition = tf.concat([condition, condition], axis=0)

        # Discrimination
        labels = tf.concat(
            [tf.zeros((batch_size, 1)), tf.ones((batch_size, 1))], axis=0
        )

        # Train the discriminator.
        with tf.GradientTape() as tape:
            disc_output, aux_output = self.discriminator(
                [combined_images, combined_condition])
            disc_d_loss = self.disc_loss_fn(labels, disc_output)
            aux_d_loss = self.aux_loss_fn(combined_condition, aux_output)
            d_loss = disc_d_loss + aux_d_loss
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # Computing D(x|y)
        d_xy = tf.math.reduce_mean(disc_output)

        ### TRAINING GENRATOR ###
        latent_noise_vector = tf.random.normal(
            shape=(batch_size, self.noise))

        # Assemble labels that say "all real images".
        misleading_labels = tf.ones((batch_size, 1))

        # Train the Generator
        with tf.GradientTape() as tape:
            fake_images = self.generator([latent_noise_vector, condition])
            disc_output, aux_output = self.discriminator(
                [fake_images, condition])
            disc_g_loss = self.disc_loss_fn(misleading_labels, disc_output)
            aux_g_loss = self.aux_loss_fn(condition, aux_output)
            g_loss = disc_g_loss + aux_g_loss
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(
            zip(grads, self.generator.trainable_weights))

        # Computing D(G(z|y))
        d_g_zy = tf.math.reduce_mean(disc_output)

        # Monitor loss and metrics.
        self.gen_loss_tracker.update_state(g_loss)
        self.disc_loss_tracker.update_state(d_loss)
        self.d_xy_tracker.update_state(d_xy)
        self.d_g_zy_tracker.update_state(d_g_zy)
        self.kl.update_state(real_images, fake_images)

        return {
            "d_loss": self.disc_loss_tracker.result(),
            "g_loss": self.gen_loss_tracker.result(),
            "D(x|y)": self.d_xy_tracker.result(),
            "D(G(z|y))": self.d_g_zy_tracker.result(),
            "KL Divergence": self.kl.result(),
        }

ACGAN Callback

To help monitor the ACGAN's loss and progress and visualise the images after every patience. We will need to make a custom tensorflow callback that will help monitor the ACGAN.

class ACGANMonitor(keras.callbacks.Callback):
    def __init__(self, num_img=20, noise=128, patience=10, vmin=0, vmax=1):
        self.num_img = num_img
        self.noise = noise
        self.patience = patience
        self.vmin = vmin
        self.vmax = vmax
        self.latent_noise_vector = tf.random.normal(
            shape=(self.num_img, self.noise))
        self.conditions = to_categorical([0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                                          0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

    def generate_plot(self):
        # Generate Images
        generated_images = self.model.generator.predict(
            [self.latent_noise_vector, self.conditions])
        # Normalise Image from [vmin, vmax] to [0, 1]
        generated_images -= self.vmin
        generated_images /= (self.vmax - self.vmin)
        row_size = int(np.ceil(self.num_img/5))
        fig = plt.figure(figsize=(10, 2*row_size), tight_layout=True)
        for i in range(self.num_img):
            ax = fig.add_subplot(row_size, 5, i+1)
            ax.imshow(generated_images[i])
            ax.set_title(class_labels[i % 10])
            ax.axis('off')
        plt.show()

    def save_weights(self, epoch=None):
        try:
            if epoch != None:
                name = 'ACGAN/generator-epoch-{}.h5'.format(epoch)
                print('Generator Checkpoint - {}'.format(name))
                self.model.generator.save_weights(
                    filepath=name,
                    save_format='h5'
                )
        except Exception as e:
            print(e)

    def on_epoch_begin(self, epoch, logs=None):
        if epoch % self.patience == 0:
            self.generate_plot()
            self.save_weights(epoch)

    def on_train_end(self, epoch, logs=None):
        self.generate_plot()
        self.save_weights('Full Train')

callbacks = [
    ACGANMonitor(num_img=20, noise=128, patience=5, vmin=-1, vmax=1),
]

ACGAN Training

After setting the ACGAN architecture, we can now run the ACGAN models to generate images. First, we need to convert the data into a tensorflow dataset and training using multiprocessing to speed up process.

tf.keras.backend.clear_session()
K.clear_session()

aux_cond_gan = AConditionalGAN(
    discriminator=create_ACGAN_discriminator(image_size=IMG_SIZE),
    generator=create_cGAN_generator(noise=NOISE),
    noise=NOISE
)

aux_cond_gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    disc_loss_fn=keras.losses.BinaryCrossentropy(),
    aux_loss_fn=keras.losses.CategoricalCrossentropy()
)

dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
dataset = dataset.shuffle(buffer_size=1024).batch(
    BATCH_SIZE, num_parallel_calls=tf.data.AUTOTUNE).prefetch(tf.data.AUTOTUNE)

aux_cond_hist = aux_cond_gan.fit(
    dataset, epochs=200, use_multiprocessing=True, workers=16, callbacks=callbacks)

1/1 [==============================] - 0s 371ms/step

Generator Checkpoint - ACGAN/generator-epoch-0.h5
Epoch 1/200
782/782 [==============================] - 75s 90ms/step - d_loss: 1.6720 - g_loss: 4.5785 - D(x|y): 0.5349 - D(G(z|y)): 0.1203 - KL Divergence: 5.3958
Epoch 2/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.3243 - g_loss: 3.9786 - D(x|y): 0.4951 - D(G(z|y)): 0.0453 - KL Divergence: 4.2473
Epoch 3/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.2607 - g_loss: 3.6651 - D(x|y): 0.5007 - D(G(z|y)): 0.0535 - KL Divergence: 5.9017
Epoch 4/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.2814 - g_loss: 3.2607 - D(x|y): 0.5010 - D(G(z|y)): 0.0685 - KL Divergence: 5.0985
Epoch 5/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2491 - g_loss: 3.5809 - D(x|y): 0.4996 - D(G(z|y)): 0.0565 - KL Divergence: 5.2626
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-5.h5
Epoch 6/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.2577 - g_loss: 3.2986 - D(x|y): 0.4998 - D(G(z|y)): 0.0740 - KL Divergence: 5.2817
Epoch 7/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2724 - g_loss: 3.1666 - D(x|y): 0.5004 - D(G(z|y)): 0.0924 - KL Divergence: 5.5191
Epoch 8/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2407 - g_loss: 3.2719 - D(x|y): 0.5004 - D(G(z|y)): 0.0859 - KL Divergence: 5.2422
Epoch 9/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2315 - g_loss: 3.2485 - D(x|y): 0.5009 - D(G(z|y)): 0.0922 - KL Divergence: 5.2537
Epoch 10/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2320 - g_loss: 3.3117 - D(x|y): 0.4987 - D(G(z|y)): 0.0934 - KL Divergence: 5.1520
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-10.h5
Epoch 11/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2672 - g_loss: 3.0352 - D(x|y): 0.4992 - D(G(z|y)): 0.1167 - KL Divergence: 5.0254
Epoch 12/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2477 - g_loss: 3.0386 - D(x|y): 0.5004 - D(G(z|y)): 0.1165 - KL Divergence: 4.9043
Epoch 13/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2460 - g_loss: 3.0714 - D(x|y): 0.5000 - D(G(z|y)): 0.1165 - KL Divergence: 4.7482
Epoch 14/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2442 - g_loss: 3.0982 - D(x|y): 0.5004 - D(G(z|y)): 0.1158 - KL Divergence: 4.7992
Epoch 15/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2402 - g_loss: 3.0818 - D(x|y): 0.5001 - D(G(z|y)): 0.1187 - KL Divergence: 4.7124
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-15.h5
Epoch 16/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2383 - g_loss: 3.1354 - D(x|y): 0.4996 - D(G(z|y)): 0.1175 - KL Divergence: 4.6767
Epoch 17/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2391 - g_loss: 3.0784 - D(x|y): 0.5007 - D(G(z|y)): 0.1222 - KL Divergence: 4.6464
Epoch 18/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2368 - g_loss: 3.1370 - D(x|y): 0.5005 - D(G(z|y)): 0.1213 - KL Divergence: 4.5927
Epoch 19/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2333 - g_loss: 3.1174 - D(x|y): 0.5012 - D(G(z|y)): 0.1266 - KL Divergence: 4.5975
Epoch 20/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2221 - g_loss: 3.1965 - D(x|y): 0.5002 - D(G(z|y)): 0.1241 - KL Divergence: 4.6999
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-20.h5
Epoch 21/200
782/782 [==============================] - 67s 86ms/step - d_loss: 0.2181 - g_loss: 3.3012 - D(x|y): 0.4998 - D(G(z|y)): 0.1191 - KL Divergence: 4.6729
Epoch 22/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.2078 - g_loss: 3.4404 - D(x|y): 0.5006 - D(G(z|y)): 0.1139 - KL Divergence: 4.6279
Epoch 23/200
782/782 [==============================] - 67s 85ms/step - d_loss: 0.1955 - g_loss: 3.5669 - D(x|y): 0.5002 - D(G(z|y)): 0.1072 - KL Divergence: 4.6501
Epoch 24/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.1867 - g_loss: 3.7490 - D(x|y): 0.5002 - D(G(z|y)): 0.1027 - KL Divergence: 4.5787
Epoch 25/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.1894 - g_loss: 3.8604 - D(x|y): 0.5000 - D(G(z|y)): 0.1006 - KL Divergence: 4.5637
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-25.h5
Epoch 26/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1712 - g_loss: 3.9667 - D(x|y): 0.5002 - D(G(z|y)): 0.0960 - KL Divergence: 4.6433
Epoch 27/200
782/782 [==============================] - 69s 89ms/step - d_loss: 0.1700 - g_loss: 3.9484 - D(x|y): 0.5003 - D(G(z|y)): 0.0964 - KL Divergence: 4.5281
Epoch 28/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1599 - g_loss: 4.2800 - D(x|y): 0.5005 - D(G(z|y)): 0.0875 - KL Divergence: 4.6995
Epoch 29/200
782/782 [==============================] - 65s 84ms/step - d_loss: 0.1652 - g_loss: 4.3234 - D(x|y): 0.5003 - D(G(z|y)): 0.0898 - KL Divergence: 4.6046
Epoch 30/200
782/782 [==============================] - 74s 94ms/step - d_loss: 0.1529 - g_loss: 4.4116 - D(x|y): 0.4997 - D(G(z|y)): 0.0877 - KL Divergence: 4.5896
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - ACGAN/generator-epoch-30.h5
Epoch 31/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1475 - g_loss: 4.6481 - D(x|y): 0.5004 - D(G(z|y)): 0.0825 - KL Divergence: 4.7073
Epoch 32/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1475 - g_loss: 4.6351 - D(x|y): 0.5005 - D(G(z|y)): 0.0820 - KL Divergence: 4.7518
Epoch 33/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1445 - g_loss: 4.7635 - D(x|y): 0.5002 - D(G(z|y)): 0.0804 - KL Divergence: 4.5958
Epoch 34/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1425 - g_loss: 4.8635 - D(x|y): 0.5009 - D(G(z|y)): 0.0792 - KL Divergence: 4.5395
Epoch 35/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.1403 - g_loss: 5.0478 - D(x|y): 0.5006 - D(G(z|y)): 0.0743 - KL Divergence: 4.5070
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-35.h5
Epoch 36/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1434 - g_loss: 5.0324 - D(x|y): 0.5002 - D(G(z|y)): 0.0781 - KL Divergence: 4.6303
Epoch 37/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1274 - g_loss: 4.9758 - D(x|y): 0.4992 - D(G(z|y)): 0.0742 - KL Divergence: 4.5329
Epoch 38/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1343 - g_loss: 5.1952 - D(x|y): 0.5005 - D(G(z|y)): 0.0743 - KL Divergence: 4.6429
Epoch 39/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1223 - g_loss: 5.4694 - D(x|y): 0.5002 - D(G(z|y)): 0.0668 - KL Divergence: 4.7246
Epoch 40/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1275 - g_loss: 5.5361 - D(x|y): 0.5000 - D(G(z|y)): 0.0675 - KL Divergence: 4.5661
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - ACGAN/generator-epoch-40.h5
Epoch 41/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1195 - g_loss: 5.6017 - D(x|y): 0.5003 - D(G(z|y)): 0.0661 - KL Divergence: 4.6437
Epoch 42/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1132 - g_loss: 5.6969 - D(x|y): 0.5001 - D(G(z|y)): 0.0646 - KL Divergence: 4.5650
Epoch 43/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1162 - g_loss: 5.8343 - D(x|y): 0.5001 - D(G(z|y)): 0.0626 - KL Divergence: 4.4984
Epoch 44/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1134 - g_loss: 5.8918 - D(x|y): 0.5001 - D(G(z|y)): 0.0618 - KL Divergence: 4.6466
Epoch 45/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.1095 - g_loss: 5.9400 - D(x|y): 0.5003 - D(G(z|y)): 0.0615 - KL Divergence: 4.6963
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-45.h5
Epoch 46/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1085 - g_loss: 5.9959 - D(x|y): 0.5003 - D(G(z|y)): 0.0598 - KL Divergence: 4.6330
Epoch 47/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1097 - g_loss: 6.1969 - D(x|y): 0.5003 - D(G(z|y)): 0.0600 - KL Divergence: 4.6365
Epoch 48/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.1056 - g_loss: 6.2659 - D(x|y): 0.5001 - D(G(z|y)): 0.0571 - KL Divergence: 4.5335
Epoch 49/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.1058 - g_loss: 6.3194 - D(x|y): 0.4998 - D(G(z|y)): 0.0575 - KL Divergence: 4.7878
Epoch 50/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1331 - g_loss: 6.0614 - D(x|y): 0.5018 - D(G(z|y)): 0.0646 - KL Divergence: 4.6140
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - ACGAN/generator-epoch-50.h5
Epoch 51/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0960 - g_loss: 6.4727 - D(x|y): 0.5001 - D(G(z|y)): 0.0542 - KL Divergence: 4.6804
Epoch 52/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0977 - g_loss: 6.5551 - D(x|y): 0.4999 - D(G(z|y)): 0.0528 - KL Divergence: 4.6576
Epoch 53/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0937 - g_loss: 6.6376 - D(x|y): 0.5005 - D(G(z|y)): 0.0509 - KL Divergence: 4.5724
Epoch 54/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0917 - g_loss: 6.8088 - D(x|y): 0.5000 - D(G(z|y)): 0.0499 - KL Divergence: 4.5355
Epoch 55/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1422 - g_loss: 6.1903 - D(x|y): 0.4999 - D(G(z|y)): 0.0664 - KL Divergence: 4.5857
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-55.h5
Epoch 56/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0917 - g_loss: 6.6402 - D(x|y): 0.5000 - D(G(z|y)): 0.0522 - KL Divergence: 4.5267
Epoch 57/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0883 - g_loss: 6.7616 - D(x|y): 0.5001 - D(G(z|y)): 0.0487 - KL Divergence: 4.6800
Epoch 58/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0893 - g_loss: 6.9044 - D(x|y): 0.5003 - D(G(z|y)): 0.0515 - KL Divergence: 4.7052
Epoch 59/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0901 - g_loss: 6.9616 - D(x|y): 0.5001 - D(G(z|y)): 0.0493 - KL Divergence: 4.6182
Epoch 60/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.1194 - g_loss: 6.7140 - D(x|y): 0.5009 - D(G(z|y)): 0.0552 - KL Divergence: 4.6978
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-60.h5
Epoch 61/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0835 - g_loss: 7.2504 - D(x|y): 0.4999 - D(G(z|y)): 0.0439 - KL Divergence: 4.5707
Epoch 62/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0819 - g_loss: 7.1562 - D(x|y): 0.4999 - D(G(z|y)): 0.0435 - KL Divergence: 4.4694
Epoch 63/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0855 - g_loss: 7.2073 - D(x|y): 0.5000 - D(G(z|y)): 0.0474 - KL Divergence: 4.6348
Epoch 64/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0849 - g_loss: 7.3213 - D(x|y): 0.5000 - D(G(z|y)): 0.0448 - KL Divergence: 4.5846
Epoch 65/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0876 - g_loss: 7.2725 - D(x|y): 0.4997 - D(G(z|y)): 0.0459 - KL Divergence: 4.5221
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-65.h5
Epoch 66/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0805 - g_loss: 7.5003 - D(x|y): 0.4998 - D(G(z|y)): 0.0416 - KL Divergence: 4.6065
Epoch 67/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0825 - g_loss: 7.4254 - D(x|y): 0.4999 - D(G(z|y)): 0.0443 - KL Divergence: 4.5317
Epoch 68/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0809 - g_loss: 7.5338 - D(x|y): 0.5003 - D(G(z|y)): 0.0418 - KL Divergence: 4.5732
Epoch 69/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0777 - g_loss: 7.5679 - D(x|y): 0.5004 - D(G(z|y)): 0.0426 - KL Divergence: 4.6603
Epoch 70/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0779 - g_loss: 7.6933 - D(x|y): 0.5002 - D(G(z|y)): 0.0406 - KL Divergence: 4.5582
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-70.h5
Epoch 71/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0770 - g_loss: 7.6865 - D(x|y): 0.5004 - D(G(z|y)): 0.0418 - KL Divergence: 4.5858
Epoch 72/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0761 - g_loss: 7.7936 - D(x|y): 0.5000 - D(G(z|y)): 0.0404 - KL Divergence: 4.6475
Epoch 73/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0745 - g_loss: 7.7917 - D(x|y): 0.5003 - D(G(z|y)): 0.0416 - KL Divergence: 4.8225
Epoch 74/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0780 - g_loss: 7.9089 - D(x|y): 0.5000 - D(G(z|y)): 0.0426 - KL Divergence: 4.6795
Epoch 75/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0726 - g_loss: 7.9694 - D(x|y): 0.5003 - D(G(z|y)): 0.0401 - KL Divergence: 4.5963
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - ACGAN/generator-epoch-75.h5
Epoch 76/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0740 - g_loss: 8.0976 - D(x|y): 0.4998 - D(G(z|y)): 0.0357 - KL Divergence: 4.6527
Epoch 77/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0689 - g_loss: 8.1082 - D(x|y): 0.5002 - D(G(z|y)): 0.0373 - KL Divergence: 4.5845
Epoch 78/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0714 - g_loss: 8.1449 - D(x|y): 0.5001 - D(G(z|y)): 0.0363 - KL Divergence: 4.5820
Epoch 79/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0698 - g_loss: 8.0619 - D(x|y): 0.4998 - D(G(z|y)): 0.0382 - KL Divergence: 4.6510
Epoch 80/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0721 - g_loss: 8.1509 - D(x|y): 0.5001 - D(G(z|y)): 0.0355 - KL Divergence: 4.5917
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-80.h5
Epoch 81/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0711 - g_loss: 8.2037 - D(x|y): 0.5000 - D(G(z|y)): 0.0371 - KL Divergence: 4.5868
Epoch 82/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0699 - g_loss: 8.2055 - D(x|y): 0.5002 - D(G(z|y)): 0.0380 - KL Divergence: 4.5820
Epoch 83/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0638 - g_loss: 8.2898 - D(x|y): 0.5000 - D(G(z|y)): 0.0346 - KL Divergence: 4.6216
Epoch 84/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0677 - g_loss: 8.4533 - D(x|y): 0.5002 - D(G(z|y)): 0.0352 - KL Divergence: 4.5663
Epoch 85/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0647 - g_loss: 8.3809 - D(x|y): 0.5004 - D(G(z|y)): 0.0364 - KL Divergence: 4.5817
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-85.h5
Epoch 86/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0667 - g_loss: 8.3906 - D(x|y): 0.5001 - D(G(z|y)): 0.0352 - KL Divergence: 4.5290
Epoch 87/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0650 - g_loss: 8.4169 - D(x|y): 0.5000 - D(G(z|y)): 0.0361 - KL Divergence: 4.4771
Epoch 88/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0645 - g_loss: 8.4398 - D(x|y): 0.5000 - D(G(z|y)): 0.0341 - KL Divergence: 4.4817
Epoch 89/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0614 - g_loss: 8.5592 - D(x|y): 0.5001 - D(G(z|y)): 0.0333 - KL Divergence: 4.5572
Epoch 90/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0607 - g_loss: 8.6394 - D(x|y): 0.5001 - D(G(z|y)): 0.0340 - KL Divergence: 4.4775
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - ACGAN/generator-epoch-90.h5
Epoch 91/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0609 - g_loss: 8.6527 - D(x|y): 0.5003 - D(G(z|y)): 0.0328 - KL Divergence: 4.4637
Epoch 92/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0624 - g_loss: 8.6129 - D(x|y): 0.4999 - D(G(z|y)): 0.0321 - KL Divergence: 4.4784
Epoch 93/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0644 - g_loss: 8.6866 - D(x|y): 0.4996 - D(G(z|y)): 0.0336 - KL Divergence: 4.6336
Epoch 94/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0624 - g_loss: 8.7968 - D(x|y): 0.5004 - D(G(z|y)): 0.0334 - KL Divergence: 4.5344
Epoch 95/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0613 - g_loss: 8.9031 - D(x|y): 0.5003 - D(G(z|y)): 0.0312 - KL Divergence: 4.4683
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - ACGAN/generator-epoch-95.h5
Epoch 96/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0655 - g_loss: 8.7648 - D(x|y): 0.5002 - D(G(z|y)): 0.0331 - KL Divergence: 4.4158
Epoch 97/200
782/782 [==============================] - 71s 90ms/step - d_loss: 0.0615 - g_loss: 8.7661 - D(x|y): 0.5006 - D(G(z|y)): 0.0324 - KL Divergence: 4.5183
Epoch 98/200
782/782 [==============================] - 70s 90ms/step - d_loss: 0.0568 - g_loss: 8.9036 - D(x|y): 0.5001 - D(G(z|y)): 0.0303 - KL Divergence: 4.4838
Epoch 99/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0563 - g_loss: 9.0012 - D(x|y): 0.5001 - D(G(z|y)): 0.0301 - KL Divergence: 4.5777
Epoch 100/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0559 - g_loss: 9.0489 - D(x|y): 0.5000 - D(G(z|y)): 0.0299 - KL Divergence: 4.6287
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - ACGAN/generator-epoch-100.h5
Epoch 101/200
782/782 [==============================] - 66s 85ms/step - d_loss: 0.0548 - g_loss: 9.0293 - D(x|y): 0.5002 - D(G(z|y)): 0.0296 - KL Divergence: 4.7024
Epoch 102/200
782/782 [==============================] - 70s 89ms/step - d_loss: 0.0576 - g_loss: 8.8470 - D(x|y): 0.5000 - D(G(z|y)): 0.0332 - KL Divergence: 4.7036
Epoch 103/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.0530 - g_loss: 9.2536 - D(x|y): 0.5001 - D(G(z|y)): 0.0279 - KL Divergence: 4.6197
Epoch 104/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0564 - g_loss: 9.0213 - D(x|y): 0.5000 - D(G(z|y)): 0.0311 - KL Divergence: 4.6464
Epoch 105/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0580 - g_loss: 9.1451 - D(x|y): 0.5003 - D(G(z|y)): 0.0304 - KL Divergence: 4.5999
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-105.h5
Epoch 106/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0535 - g_loss: 9.1489 - D(x|y): 0.5002 - D(G(z|y)): 0.0283 - KL Divergence: 4.5447
Epoch 107/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0540 - g_loss: 9.1239 - D(x|y): 0.5000 - D(G(z|y)): 0.0295 - KL Divergence: 4.4865
Epoch 108/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0548 - g_loss: 9.2245 - D(x|y): 0.5001 - D(G(z|y)): 0.0312 - KL Divergence: 4.5559
Epoch 109/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0568 - g_loss: 9.1778 - D(x|y): 0.4995 - D(G(z|y)): 0.0286 - KL Divergence: 4.5003
Epoch 110/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0568 - g_loss: 9.2491 - D(x|y): 0.5002 - D(G(z|y)): 0.0302 - KL Divergence: 4.6893
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - ACGAN/generator-epoch-110.h5
Epoch 111/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0567 - g_loss: 9.3953 - D(x|y): 0.5002 - D(G(z|y)): 0.0273 - KL Divergence: 4.6287
Epoch 112/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0528 - g_loss: 9.3565 - D(x|y): 0.5003 - D(G(z|y)): 0.0282 - KL Divergence: 4.7043
Epoch 113/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0526 - g_loss: 9.4539 - D(x|y): 0.5001 - D(G(z|y)): 0.0285 - KL Divergence: 4.5758
Epoch 114/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0507 - g_loss: 9.5205 - D(x|y): 0.5001 - D(G(z|y)): 0.0257 - KL Divergence: 4.4659
Epoch 115/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0531 - g_loss: 9.5134 - D(x|y): 0.5001 - D(G(z|y)): 0.0271 - KL Divergence: 4.5143
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-115.h5
Epoch 116/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0511 - g_loss: 9.5440 - D(x|y): 0.4998 - D(G(z|y)): 0.0289 - KL Divergence: 4.3214
Epoch 117/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0507 - g_loss: 9.6531 - D(x|y): 0.5002 - D(G(z|y)): 0.0258 - KL Divergence: 4.4813
Epoch 118/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0499 - g_loss: 9.5880 - D(x|y): 0.5003 - D(G(z|y)): 0.0257 - KL Divergence: 4.4493
Epoch 119/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0501 - g_loss: 9.5242 - D(x|y): 0.5000 - D(G(z|y)): 0.0281 - KL Divergence: 4.3927
Epoch 120/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0471 - g_loss: 9.4812 - D(x|y): 0.5002 - D(G(z|y)): 0.0273 - KL Divergence: 4.4178
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-120.h5
Epoch 121/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0497 - g_loss: 9.5106 - D(x|y): 0.5000 - D(G(z|y)): 0.0252 - KL Divergence: 4.4753
Epoch 122/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0535 - g_loss: 9.4896 - D(x|y): 0.5005 - D(G(z|y)): 0.0269 - KL Divergence: 4.4139
Epoch 123/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0487 - g_loss: 9.4986 - D(x|y): 0.5000 - D(G(z|y)): 0.0264 - KL Divergence: 4.5183
Epoch 124/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0470 - g_loss: 9.5799 - D(x|y): 0.5001 - D(G(z|y)): 0.0269 - KL Divergence: 4.5259
Epoch 125/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0479 - g_loss: 9.5899 - D(x|y): 0.5003 - D(G(z|y)): 0.0250 - KL Divergence: 4.5949
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-125.h5
Epoch 126/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0484 - g_loss: 9.5200 - D(x|y): 0.5002 - D(G(z|y)): 0.0241 - KL Divergence: 4.6112
Epoch 127/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0461 - g_loss: 9.7063 - D(x|y): 0.5001 - D(G(z|y)): 0.0243 - KL Divergence: 4.5735
Epoch 128/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0483 - g_loss: 9.6515 - D(x|y): 0.5003 - D(G(z|y)): 0.0269 - KL Divergence: 4.6244
Epoch 129/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0472 - g_loss: 9.6410 - D(x|y): 0.4998 - D(G(z|y)): 0.0252 - KL Divergence: 4.5093
Epoch 130/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0511 - g_loss: 9.7725 - D(x|y): 0.5000 - D(G(z|y)): 0.0247 - KL Divergence: 4.4478
1/1 [==============================] - 0s 34ms/step

Generator Checkpoint - ACGAN/generator-epoch-130.h5
Epoch 131/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0483 - g_loss: 9.8853 - D(x|y): 0.5004 - D(G(z|y)): 0.0250 - KL Divergence: 4.5073
Epoch 132/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0472 - g_loss: 9.9363 - D(x|y): 0.4998 - D(G(z|y)): 0.0233 - KL Divergence: 4.5257
Epoch 133/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0465 - g_loss: 9.9072 - D(x|y): 0.5002 - D(G(z|y)): 0.0241 - KL Divergence: 4.4540
Epoch 134/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0485 - g_loss: 9.9278 - D(x|y): 0.4999 - D(G(z|y)): 0.0233 - KL Divergence: 4.5996
Epoch 135/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0433 - g_loss: 10.0673 - D(x|y): 0.5000 - D(G(z|y)): 0.0235 - KL Divergence: 4.4603
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-135.h5
Epoch 136/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0437 - g_loss: 10.0597 - D(x|y): 0.4997 - D(G(z|y)): 0.0230 - KL Divergence: 4.6248
Epoch 137/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0448 - g_loss: 10.1266 - D(x|y): 0.5001 - D(G(z|y)): 0.0229 - KL Divergence: 4.6326
Epoch 138/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0414 - g_loss: 10.1001 - D(x|y): 0.5000 - D(G(z|y)): 0.0220 - KL Divergence: 4.8064
Epoch 139/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0447 - g_loss: 9.9543 - D(x|y): 0.5001 - D(G(z|y)): 0.0229 - KL Divergence: 4.6194
Epoch 140/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0416 - g_loss: 9.9515 - D(x|y): 0.5005 - D(G(z|y)): 0.0237 - KL Divergence: 4.5918
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - ACGAN/generator-epoch-140.h5
Epoch 141/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0439 - g_loss: 10.1954 - D(x|y): 0.4999 - D(G(z|y)): 0.0224 - KL Divergence: 4.6564
Epoch 142/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0445 - g_loss: 10.1081 - D(x|y): 0.4996 - D(G(z|y)): 0.0244 - KL Divergence: 4.4255
Epoch 143/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0427 - g_loss: 10.2899 - D(x|y): 0.5003 - D(G(z|y)): 0.0226 - KL Divergence: 4.5401
Epoch 144/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0415 - g_loss: 10.2155 - D(x|y): 0.4999 - D(G(z|y)): 0.0209 - KL Divergence: 4.6224
Epoch 145/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0396 - g_loss: 10.2334 - D(x|y): 0.5002 - D(G(z|y)): 0.0228 - KL Divergence: 4.6463
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - ACGAN/generator-epoch-145.h5
Epoch 146/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0412 - g_loss: 10.2631 - D(x|y): 0.5004 - D(G(z|y)): 0.0221 - KL Divergence: 4.6867
Epoch 147/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0410 - g_loss: 10.3057 - D(x|y): 0.5001 - D(G(z|y)): 0.0204 - KL Divergence: 4.6379
Epoch 148/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0419 - g_loss: 10.2060 - D(x|y): 0.5000 - D(G(z|y)): 0.0221 - KL Divergence: 4.5885
Epoch 149/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0399 - g_loss: 10.3822 - D(x|y): 0.5003 - D(G(z|y)): 0.0225 - KL Divergence: 4.7194
Epoch 150/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0380 - g_loss: 10.3306 - D(x|y): 0.5002 - D(G(z|y)): 0.0211 - KL Divergence: 4.5707
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-150.h5
Epoch 151/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0382 - g_loss: 10.4163 - D(x|y): 0.5002 - D(G(z|y)): 0.0217 - KL Divergence: 4.4491
Epoch 152/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0388 - g_loss: 10.5753 - D(x|y): 0.5002 - D(G(z|y)): 0.0202 - KL Divergence: 4.5556
Epoch 153/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0407 - g_loss: 10.4754 - D(x|y): 0.5001 - D(G(z|y)): 0.0213 - KL Divergence: 4.4323
Epoch 154/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0414 - g_loss: 10.4811 - D(x|y): 0.5002 - D(G(z|y)): 0.0191 - KL Divergence: 4.4706
Epoch 155/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0360 - g_loss: 10.4229 - D(x|y): 0.5002 - D(G(z|y)): 0.0208 - KL Divergence: 4.5416
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-155.h5
Epoch 156/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0391 - g_loss: 10.5886 - D(x|y): 0.5002 - D(G(z|y)): 0.0212 - KL Divergence: 4.5294
Epoch 157/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0356 - g_loss: 10.5692 - D(x|y): 0.5001 - D(G(z|y)): 0.0188 - KL Divergence: 4.6173
Epoch 158/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0344 - g_loss: 10.5680 - D(x|y): 0.5000 - D(G(z|y)): 0.0204 - KL Divergence: 4.6693
Epoch 159/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0381 - g_loss: 10.3157 - D(x|y): 0.5002 - D(G(z|y)): 0.0209 - KL Divergence: 4.5108
Epoch 160/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0389 - g_loss: 10.5065 - D(x|y): 0.5002 - D(G(z|y)): 0.0211 - KL Divergence: 4.4974
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-160.h5
Epoch 161/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0389 - g_loss: 10.3096 - D(x|y): 0.5001 - D(G(z|y)): 0.0197 - KL Divergence: 4.5462
Epoch 162/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0380 - g_loss: 10.5626 - D(x|y): 0.5002 - D(G(z|y)): 0.0218 - KL Divergence: 4.4762
Epoch 163/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0390 - g_loss: 10.5609 - D(x|y): 0.5001 - D(G(z|y)): 0.0208 - KL Divergence: 4.5961
Epoch 164/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0377 - g_loss: 10.6561 - D(x|y): 0.5002 - D(G(z|y)): 0.0205 - KL Divergence: 4.4667
Epoch 165/200
782/782 [==============================] - 65s 82ms/step - d_loss: 0.0391 - g_loss: 10.3624 - D(x|y): 0.5002 - D(G(z|y)): 0.0193 - KL Divergence: 4.5474
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-165.h5
Epoch 166/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0376 - g_loss: 10.8100 - D(x|y): 0.4998 - D(G(z|y)): 0.0191 - KL Divergence: 4.3804
Epoch 167/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0341 - g_loss: 10.7767 - D(x|y): 0.5003 - D(G(z|y)): 0.0193 - KL Divergence: 4.4211
Epoch 168/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0344 - g_loss: 10.5924 - D(x|y): 0.5000 - D(G(z|y)): 0.0191 - KL Divergence: 4.5924
Epoch 169/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0372 - g_loss: 10.6332 - D(x|y): 0.5001 - D(G(z|y)): 0.0194 - KL Divergence: 4.5628
Epoch 170/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0344 - g_loss: 10.6284 - D(x|y): 0.5001 - D(G(z|y)): 0.0184 - KL Divergence: 4.5610
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - ACGAN/generator-epoch-170.h5
Epoch 171/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0389 - g_loss: 10.6685 - D(x|y): 0.5001 - D(G(z|y)): 0.0200 - KL Divergence: 4.5662
Epoch 172/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0385 - g_loss: 10.4595 - D(x|y): 0.5003 - D(G(z|y)): 0.0204 - KL Divergence: 4.6869
Epoch 173/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0330 - g_loss: 10.7177 - D(x|y): 0.5002 - D(G(z|y)): 0.0178 - KL Divergence: 4.5463
Epoch 174/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0348 - g_loss: 10.8425 - D(x|y): 0.5001 - D(G(z|y)): 0.0186 - KL Divergence: 4.5929
Epoch 175/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0354 - g_loss: 10.8588 - D(x|y): 0.5000 - D(G(z|y)): 0.0183 - KL Divergence: 4.4169
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - ACGAN/generator-epoch-175.h5
Epoch 176/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0356 - g_loss: 10.7409 - D(x|y): 0.5000 - D(G(z|y)): 0.0180 - KL Divergence: 4.6203
Epoch 177/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0345 - g_loss: 10.6794 - D(x|y): 0.4999 - D(G(z|y)): 0.0189 - KL Divergence: 4.6679
Epoch 178/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0379 - g_loss: 10.7066 - D(x|y): 0.4997 - D(G(z|y)): 0.0199 - KL Divergence: 4.6324
Epoch 179/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0337 - g_loss: 11.0204 - D(x|y): 0.5003 - D(G(z|y)): 0.0171 - KL Divergence: 4.4794
Epoch 180/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0362 - g_loss: 10.9560 - D(x|y): 0.4999 - D(G(z|y)): 0.0187 - KL Divergence: 4.5774
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - ACGAN/generator-epoch-180.h5
Epoch 181/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0348 - g_loss: 11.0296 - D(x|y): 0.5002 - D(G(z|y)): 0.0182 - KL Divergence: 4.6194
Epoch 182/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0311 - g_loss: 11.1445 - D(x|y): 0.5002 - D(G(z|y)): 0.0176 - KL Divergence: 4.5359
Epoch 183/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.0313 - g_loss: 10.9244 - D(x|y): 0.5002 - D(G(z|y)): 0.0177 - KL Divergence: 4.4815
Epoch 184/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0388 - g_loss: 10.6597 - D(x|y): 0.4999 - D(G(z|y)): 0.0186 - KL Divergence: 4.6581
Epoch 185/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0353 - g_loss: 10.8569 - D(x|y): 0.5000 - D(G(z|y)): 0.0179 - KL Divergence: 4.5563
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - ACGAN/generator-epoch-185.h5
Epoch 186/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0356 - g_loss: 10.5855 - D(x|y): 0.5002 - D(G(z|y)): 0.0170 - KL Divergence: 4.6060
Epoch 187/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0414 - g_loss: 10.3453 - D(x|y): 0.5004 - D(G(z|y)): 0.0181 - KL Divergence: 4.4798
Epoch 188/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0391 - g_loss: 10.6659 - D(x|y): 0.5001 - D(G(z|y)): 0.0177 - KL Divergence: 4.7812
Epoch 189/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0378 - g_loss: 10.2462 - D(x|y): 0.5004 - D(G(z|y)): 0.0174 - KL Divergence: 4.6804
Epoch 190/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0401 - g_loss: 9.8650 - D(x|y): 0.5007 - D(G(z|y)): 0.0180 - KL Divergence: 4.9386
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - ACGAN/generator-epoch-190.h5
Epoch 191/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0374 - g_loss: 9.9178 - D(x|y): 0.5006 - D(G(z|y)): 0.0188 - KL Divergence: 4.7206
Epoch 192/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0426 - g_loss: 10.2274 - D(x|y): 0.5004 - D(G(z|y)): 0.0184 - KL Divergence: 4.7145
Epoch 193/200
782/782 [==============================] - 68s 88ms/step - d_loss: 0.0405 - g_loss: 9.9893 - D(x|y): 0.4998 - D(G(z|y)): 0.0176 - KL Divergence: 4.2209
Epoch 194/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0451 - g_loss: 10.1866 - D(x|y): 0.5003 - D(G(z|y)): 0.0184 - KL Divergence: 4.1549
Epoch 195/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0426 - g_loss: 9.9711 - D(x|y): 0.5004 - D(G(z|y)): 0.0198 - KL Divergence: 4.4706
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - ACGAN/generator-epoch-195.h5
Epoch 196/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0433 - g_loss: 9.8577 - D(x|y): 0.5003 - D(G(z|y)): 0.0206 - KL Divergence: 4.8519
Epoch 197/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.0470 - g_loss: 9.9979 - D(x|y): 0.5002 - D(G(z|y)): 0.0196 - KL Divergence: 4.9123
Epoch 198/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0490 - g_loss: 9.8164 - D(x|y): 0.5001 - D(G(z|y)): 0.0194 - KL Divergence: 4.2961
Epoch 199/200
782/782 [==============================] - 69s 88ms/step - d_loss: 0.0519 - g_loss: 9.6917 - D(x|y): 0.4999 - D(G(z|y)): 0.0216 - KL Divergence: 4.1689
Epoch 200/200
782/782 [==============================] - 68s 87ms/step - d_loss: 0.0477 - g_loss: 9.4337 - D(x|y): 0.5003 - D(G(z|y)): 0.0199 - KL Divergence: 4.4272
1/1 [==============================] - 0s 32ms/step

Generator Checkpoint - ACGAN/generator-epoch-Full Train.h5

Observations

Similarly to the cGAN, the ACGAN has a issue with same images generating again using the same label and noise. However, this problem only occurred during the later part of the training around 185 - 190. We also notice that the images around 126, there are noise inputs are showing slightly

ACGAN Evaluation

As mentioned previously we will be using a quantitative and manual evaluation of the ACGAN.

We will start off with a quantitative analysis so that we can find the best model to generate the images for manual evaluation.

# story history object into dataframe
aux_cond_hist_df = pd.DataFrame(aux_cond_hist.history)

# using pandas dataframe to plot out learning curve
fig, (ax1, ax2, ax3) = plt.subplots(
    1, 3, figsize=(24, 8), tight_layout=True)
aux_cond_hist_df.loc[:, ["d_loss", 'g_loss']].plot(
    ax=ax1, title=r'Learning Curve of Loss Function CE')
aux_cond_hist_df.loc[:, ['D(x|y)', 'D(G(z|y))']].plot(
    ax=ax2, title=r'Learning Curve of Discriminator')
aux_cond_hist_df.loc[:, 'KL Divergence'].plot(
    ax=ax3, title=r'Learning Curve of Probabilistic Similarity')
plt.show()

Observations

We note that as the model has collapsed early around 185 epochs. The lowest KL divergence is 115 epochs.

# Loading Weights for best Generator
saved_weights = 'ACGAN\generator-epoch-115.h5'
aux_cond_gan.generator.load_weights(saved_weights)
aux_cond_gan.generator.summary()

Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_16 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_14 (Dense)               (None, 2048)         284672      ['concatenate_16[0][0]']         
                                                                                                  
 batch_normalization_150 (Batch  (None, 2048)        8192        ['dense_14[0][0]']               
 Normalization)                                                                                   
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_16 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_14 (Dense)               (None, 2048)         284672      ['concatenate_16[0][0]']         
                                                                                                  
 batch_normalization_150 (Batch  (None, 2048)        8192        ['dense_14[0][0]']               
 Normalization)                                                                                   
                                                                                                  
 leaky_re_lu_40 (LeakyReLU)     (None, 2048)         0           ['batch_normalization_150[0][0]']
                                                                                                  
 reshape_5 (Reshape)            (None, 2, 2, 512)    0           ['leaky_re_lu_40[0][0]']         
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2754752     ['reshape_5[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,050,691
Trainable params: 3,045,699
Non-trainable params: 4,992
__________________________________________________________________________________________________

n = 10000

# generating labels
labels = np.random.randint(low=0, high=10, size=n)
one_hot_labels = to_categorical(labels)

# Generating 1000 Synthetic Images
random_noise = tf.random.normal(shape=(n, NOISE))

synthetic_images = aux_cond_gan.generator.predict(
    [random_noise, one_hot_labels])
print("Latent Vector Dim: {}\tGenerated Images Dim: {}".format(
    random_noise.shape, synthetic_images.shape))

# Scaling back to [0, 1]
synthetic_images -= -1
synthetic_images /= (1 - (-1))

# Display 25 randomly sampled images
fig = plt.figure(figsize=(10, 10), tight_layout=True)
for i in range(25):
    rand_idx = np.random.randint(0, len(synthetic_images))
    ax = fig.add_subplot(5, 5, i+1)
    ax.imshow(synthetic_images[rand_idx])
    ax.set_title(class_labels[labels[rand_idx]])
    ax.axis('off')
plt.show()

313/313 [==============================] - 4s 11ms/step
Latent Vector Dim: (10000, 128)	Generated Images Dim: (10000, 32, 32, 3)

Observations

We can see the generator being able to more clearly draw out the vehicles as well as same of the animals like deer and horse. However, some of the animals like cat and dog etc are still very distorted. Some changes should be made to the architecture like introducing spectral normalisation or regularisation to allow the model to generalise and learn features faster.

aux_cond_gan_fid_class = GAN_FID(batch_size=512, noise=128,
                                 sample_size=10000, buffer_size=1024, labels=True)
aux_cond_gan_fid_class.fit(
    generator=aux_cond_gan.generator, train_data=x_train)
fid_score = aux_cond_gan_fid_class.evaluate()

Computing Real Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 6s 122ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 126ms/step
16/16 [==============================] - 2s 125ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 126ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 129ms/step
Computing Generated Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 134ms/step
16/16 [==============================] - 2s 133ms/step
Computed Embeddings	Real Images Embedding Shape: (10240, 2048)	Generated Images Embedding Shape: (10240, 2048)
The computed FID score is: 88.7613090961658

Observations

We note that the FID score is now lower and drop from 94.536161025068 to 88.7613090961658. This suggest that more needs to be done to allow the FID score to be decreased.

Model Selection

Based on the different model architecture that we have tested, we can see the more complex ACGAN performed better as it is able to have clear images of more classes than the cGAN and DCGAN. It is also able to have a lower FID score compared to both cGAN and DCGAN. However, as FID score is not a good indication for the model generation we will be also improving the cGAN model.

Model Improvements

ACGAN

New ACGAN Discriminator

As mentioned in the DCGAN discriminator creation, we will be using Gaussian Weights layers for the new ACGAN model.

Gaussian weights will reduce the instability of training [mean = 0, std = 0.02].

# function to create discriminator model
def create_improve_ACGAN_discriminator(image_size):
    # input image
    img_input = Input(shape=(image_size), name='Image_Input')
    weights_init = RandomNormal(mean=0, stddev=0.02)

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    base_discriminator = Sequential([
        Conv2D(64, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(128, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(256, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(512, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
    ], name='Base_Discriminator')
    discriminator = base_discriminator(img_input)

    discriminator = GlobalAveragePooling2D()(discriminator)

    # Concatenate - combine with conditions y
    merged_layer = Concatenate()([discriminator, conditions])
    discriminator = Dense(512, activation='relu')(merged_layer)

    # Output
    Disc_Output_Layer = Dense(1, activation='sigmoid',
                              name='Disc_Output_Layer')(discriminator)
    Aux_Output_Layer = Dense(10, activation='softmax',
                             name='Aux_Output_Layer')(discriminator)

    discriminator = Model(inputs=[img_input, conditions],
                          outputs=[Disc_Output_Layer, Aux_Output_Layer], name='discriminator_ACGAN')
    return discriminator


create_improve_ACGAN_discriminator(image_size=IMG_SIZE).summary()

c:\Users\Soh Hong Yu\anaconda3\envs\gpu_env\lib\site-packages\keras\initializers\initializers_v2.py:120: UserWarning: The initializer RandomNormal is unseeded and being called multiple times, which will return identical values  each time (even if the initializer is unseeded). Please update your code to provide a seed to the initializer, or avoid using the same initalizer instance more than once.
  warnings.warn(

Model: "discriminator_ACGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
                                                                                                  
 global_average_pooling2d (Glob  (None, 512)         0           ['Base_Discriminator[0][0]']     
 alAveragePooling2D)                                                                              
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate (Concatenate)      (None, 522)          0           ['global_average_pooling2d[0][0]'
                                                                 , 'Conditions_y[0][0]']          
                                                                                                  
 dense (Dense)                  (None, 512)          267776      ['concatenate[0][0]']            
                                                                                                  
 Disc_Output_Layer (Dense)      (None, 1)            513         ['dense[0][0]']                  
                                                                                                  
 Aux_Output_Layer (Dense)       (None, 10)           5130        ['dense[0][0]']                  
                                                                                                  
==================================================================================================
Total params: 3,033,803
Trainable params: 3,031,883
Non-trainable params: 1,920
__________________________________________________________________________________________________

plot_model(create_improve_ACGAN_discriminator(image_size=IMG_SIZE), to_file="images/models/New ACGAN_discriminator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

New ACGAN Generator

Similarly, we will use Spectral Normalisation, Gaussian weights and PReLU for the generator. This will allow the generator to adapt better and learn better overall compared to the LeakyReLU used previously.

# function to create generator model
def create_improve_ACGAN_generator(noise):

    # gaussian weights initialization
    weights_init = RandomNormal(mean=0, stddev=0.02)

    # latent noise vector z
    z = Input(shape=(noise,), name="Latent_Noise_Vector_z")

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    # Generator network
    merged_layer = Concatenate()([z, conditions])

    # FC: 2x2x512
    generator = Dense(2*2*512, activation='relu')(merged_layer)
    generator = BatchNormalization(momentum=0.8)(generator)
    generator = PReLU()(generator)
    generator = Reshape((2, 2, 512))(generator)

    base_generator = Sequential([
        # Conv 1: 4x4x256
        SpectralNormalization(Conv2DTranspose(256, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
        # Conv 2: 8x8x128
        SpectralNormalization(Conv2DTranspose(128, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
        # Conv 3: 16x16x64
        SpectralNormalization(Conv2DTranspose(64, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
    ], name='Base_Generator')
    generator = base_generator(generator)

    # Conv 4: 32x32x3
    generator = Conv2DTranspose(3, kernel_size=4, strides=2, padding='same',
                                activation='tanh', name='Output_Layer')(generator)

    generator = Model(inputs=[z, conditions],
                      outputs=generator,  name='generator_ACGAN')
    return generator


create_improve_ACGAN_generator(noise=NOISE).summary()

Model: "generator_ACGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_1 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_1 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_1 (Dense)                (None, 2048)         284672      ['concatenate_1[0][0]']          
                                                                                                  
 batch_normalization_4 (BatchNo  (None, 2048)        8192        ['dense_1[0][0]']                
 rmalization)                                                                                     
                                                                                                  
 p_re_lu (PReLU)                (None, 2048)         2048        ['batch_normalization_4[0][0]']  
                                                                                                  
 reshape (Reshape)              (None, 2, 2, 512)    0           ['p_re_lu[0][0]']                
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape[0][0]']                
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

plot_model(create_improve_ACGAN_generator(noise=NOISE), to_file="images/models/New ACGAN_generator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

New ACGAN Callback

class NewACGANMonitor(keras.callbacks.Callback):
    def __init__(self, num_img=20, noise=128, patience=10, vmin=0, vmax=1):
        self.num_img = num_img
        self.noise = noise
        self.patience = patience
        self.vmin = vmin
        self.vmax = vmax
        self.latent_noise_vector = tf.random.normal(
            shape=(self.num_img, self.noise))
        self.conditions = to_categorical([0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                                          0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

    def generate_plot(self):
        # Generate Images
        generated_images = self.model.generator.predict(
            [self.latent_noise_vector, self.conditions])
        # Normalise Image from [vmin, vmax] to [0, 1]
        generated_images -= self.vmin
        generated_images /= (self.vmax - self.vmin)
        row_size = int(np.ceil(self.num_img/5))
        fig = plt.figure(figsize=(10, 2*row_size), tight_layout=True)
        for i in range(self.num_img):
            ax = fig.add_subplot(row_size, 5, i+1)
            ax.imshow(generated_images[i])
            ax.set_title(class_labels[i % 10])
            ax.axis('off')
        plt.show()

    def save_weights(self, epoch=None):
        try:
            if epoch != None:
                name = 'New ACGAN/generator-epoch-{}.h5'.format(epoch)
                print('Generator Checkpoint - {}'.format(name))
                self.model.generator.save_weights(
                    filepath=name,
                    save_format='h5'
                )
        except Exception as e:
            print(e)

    def on_epoch_begin(self, epoch, logs=None):
        if epoch % self.patience == 0:
            self.generate_plot()
            self.save_weights(epoch)

    def on_train_end(self, epoch, logs=None):
        self.generate_plot()
        self.save_weights('Full Train')

callbacks = [
    NewACGANMonitor(num_img=20, noise=128, patience=5, vmin=-1, vmax=1),
]

New ACGAN Training

class AConditionalGAN(tf.keras.Model):
    def __init__(self, discriminator, generator, noise):
        super(AConditionalGAN, self).__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.noise = noise
        self.gen_loss_tracker = tf.keras.metrics.Mean(name="generator_loss")
        self.disc_loss_tracker = tf.keras.metrics.Mean(
            name="discriminator_loss")
        self.d_xy_tracker = tf.keras.metrics.Mean(name='Mean D(x|y)')
        self.d_g_zy_tracker = tf.keras.metrics.Mean(name='Mean D(G(z|y))')
        self.kl = tf.keras.metrics.KLDivergence()

    def compile(self, d_optimizer, g_optimizer, disc_loss_fn, aux_loss_fn):
        super(AConditionalGAN, self).compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.disc_loss_fn = disc_loss_fn
        self.aux_loss_fn = aux_loss_fn

    def train_step(self, data):
        ### TRAINING DISCRIMINATOR ###
        # Unpack the data.
        real_images, condition = data

        # Sample for latent noise vector z
        batch_size = tf.shape(real_images)[0]
        latent_noise_vector = tf.random.normal(
            shape=(batch_size, self.noise))

        # Maps the noise latent vector and labels to generate fake images.
        generated_images = self.generator([latent_noise_vector, condition])

        # Combine with real images
        combined_images = tf.concat([generated_images, real_images], axis=0)
        combined_condition = tf.concat([condition, condition], axis=0)

        # Discrimination
        labels = tf.concat(
            [tf.zeros((batch_size, 1)), tf.ones((batch_size, 1))], axis=0
        )

        # Train the discriminator.
        with tf.GradientTape() as tape:
            disc_output, aux_output = self.discriminator(
                [combined_images, combined_condition])
            disc_d_loss = self.disc_loss_fn(labels, disc_output)
            aux_d_loss = self.aux_loss_fn(combined_condition, aux_output)
            d_loss = disc_d_loss + aux_d_loss
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # Computing D(x|y)
        d_xy = tf.math.reduce_mean(disc_output)

        ### TRAINING GENRATOR ###
        latent_noise_vector = tf.random.normal(
            shape=(batch_size, self.noise))

        # Assemble labels that say "all real images".
        misleading_labels = tf.ones((batch_size, 1))

        # Train the Generator
        with tf.GradientTape() as tape:
            fake_images = self.generator([latent_noise_vector, condition])
            disc_output, aux_output = self.discriminator(
                [fake_images, condition])
            disc_g_loss = self.disc_loss_fn(misleading_labels, disc_output)
            aux_g_loss = self.aux_loss_fn(condition, aux_output)
            g_loss = disc_g_loss + aux_g_loss
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(
            zip(grads, self.generator.trainable_weights))

        # Computing D(G(z|y))
        d_g_zy = tf.math.reduce_mean(disc_output)

        # Monitor loss and metrics.
        self.gen_loss_tracker.update_state(g_loss)
        self.disc_loss_tracker.update_state(d_loss)
        self.d_xy_tracker.update_state(d_xy)
        self.d_g_zy_tracker.update_state(d_g_zy)
        self.kl.update_state(real_images, fake_images)

        return {
            "d_loss": self.disc_loss_tracker.result(),
            "g_loss": self.gen_loss_tracker.result(),
            "D(x|y)": self.d_xy_tracker.result(),
            "D(G(z|y))": self.d_g_zy_tracker.result(),
            "KL Divergence": self.kl.result(),
        }

tf.keras.backend.clear_session()
K.clear_session()

improve_aux_cond_gan = AConditionalGAN(
    discriminator=create_improve_ACGAN_discriminator(image_size=IMG_SIZE),
    generator=create_improve_ACGAN_generator(noise=NOISE),
    noise=NOISE
)

improve_aux_cond_gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    disc_loss_fn=keras.losses.BinaryCrossentropy(),
    aux_loss_fn=keras.losses.CategoricalCrossentropy()
)

dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
dataset = dataset.shuffle(buffer_size=1024).batch(
    BATCH_SIZE, num_parallel_calls=tf.data.AUTOTUNE).prefetch(tf.data.AUTOTUNE)

improve_aux_cond_hist = improve_aux_cond_gan.fit(
    dataset, epochs=200, use_multiprocessing=True, workers=16, callbacks=callbacks)

1/1 [==============================] - 4s 4s/step

Generator Checkpoint - New ACGAN/generator-epoch-0.h5
Epoch 1/200
782/782 [==============================] - 109s 87ms/step - d_loss: 1.6439 - g_loss: 4.4664 - D(x|y): 0.5262 - D(G(z|y)): 0.1249 - KL Divergence: 5.0581
Epoch 2/200
782/782 [==============================] - 88s 112ms/step - d_loss: 0.3334 - g_loss: 3.3972 - D(x|y): 0.4955 - D(G(z|y)): 0.0696 - KL Divergence: 5.6292
Epoch 3/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.3269 - g_loss: 3.3939 - D(x|y): 0.4955 - D(G(z|y)): 0.0747 - KL Divergence: 4.8018
Epoch 4/200
782/782 [==============================] - 96s 123ms/step - d_loss: 0.2873 - g_loss: 2.9727 - D(x|y): 0.5010 - D(G(z|y)): 0.0848 - KL Divergence: 5.2614
Epoch 5/200
782/782 [==============================] - 312s 399ms/step - d_loss: 0.2686 - g_loss: 3.1902 - D(x|y): 0.5025 - D(G(z|y)): 0.0723 - KL Divergence: 4.7380
1/1 [==============================] - 0s 40ms/step

Generator Checkpoint - New ACGAN/generator-epoch-5.h5
Epoch 6/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.2766 - g_loss: 3.1532 - D(x|y): 0.5018 - D(G(z|y)): 0.0753 - KL Divergence: 5.1060
Epoch 7/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.2999 - g_loss: 2.9464 - D(x|y): 0.5010 - D(G(z|y)): 0.1009 - KL Divergence: 5.0076
Epoch 8/200
782/782 [==============================] - 77s 99ms/step - d_loss: 0.2905 - g_loss: 2.9631 - D(x|y): 0.4994 - D(G(z|y)): 0.1016 - KL Divergence: 4.5838
Epoch 9/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.2880 - g_loss: 2.9228 - D(x|y): 0.5009 - D(G(z|y)): 0.1095 - KL Divergence: 4.5984
Epoch 10/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.2843 - g_loss: 2.8786 - D(x|y): 0.5000 - D(G(z|y)): 0.1158 - KL Divergence: 4.6413
1/1 [==============================] - 0s 41ms/step

Generator Checkpoint - New ACGAN/generator-epoch-10.h5
Epoch 11/200
782/782 [==============================] - 73s 93ms/step - d_loss: 0.2865 - g_loss: 2.8325 - D(x|y): 0.5019 - D(G(z|y)): 0.1232 - KL Divergence: 4.5089
Epoch 12/200
782/782 [==============================] - 72s 93ms/step - d_loss: 0.3141 - g_loss: 2.6075 - D(x|y): 0.4996 - D(G(z|y)): 0.1471 - KL Divergence: 4.4697
Epoch 13/200
782/782 [==============================] - 71s 91ms/step - d_loss: 0.3046 - g_loss: 2.6422 - D(x|y): 0.5013 - D(G(z|y)): 0.1438 - KL Divergence: 4.6338
Epoch 14/200
782/782 [==============================] - 196s 250ms/step - d_loss: 0.3209 - g_loss: 2.5556 - D(x|y): 0.5005 - D(G(z|y)): 0.1546 - KL Divergence: 4.5383
Epoch 15/200
782/782 [==============================] - 72s 92ms/step - d_loss: 0.3140 - g_loss: 2.5321 - D(x|y): 0.5002 - D(G(z|y)): 0.1566 - KL Divergence: 4.5920
1/1 [==============================] - 0s 46ms/step

Generator Checkpoint - New ACGAN/generator-epoch-15.h5
Epoch 16/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.3165 - g_loss: 2.4437 - D(x|y): 0.5009 - D(G(z|y)): 0.1664 - KL Divergence: 4.6190
Epoch 17/200
782/782 [==============================] - 78s 99ms/step - d_loss: 0.3044 - g_loss: 2.5265 - D(x|y): 0.5005 - D(G(z|y)): 0.1592 - KL Divergence: 4.5491
Epoch 18/200
782/782 [==============================] - 92s 117ms/step - d_loss: 0.3748 - g_loss: 2.0873 - D(x|y): 0.4971 - D(G(z|y)): 0.2152 - KL Divergence: 4.6130
Epoch 19/200
782/782 [==============================] - 74s 95ms/step - d_loss: 0.3095 - g_loss: 2.5011 - D(x|y): 0.4999 - D(G(z|y)): 0.1647 - KL Divergence: 4.5178
Epoch 20/200
782/782 [==============================] - 152s 194ms/step - d_loss: 0.2898 - g_loss: 2.5939 - D(x|y): 0.5002 - D(G(z|y)): 0.1593 - KL Divergence: 4.6552
1/1 [==============================] - 0s 98ms/step

Generator Checkpoint - New ACGAN/generator-epoch-20.h5
Epoch 21/200
782/782 [==============================] - 75s 95ms/step - d_loss: 0.2879 - g_loss: 2.6576 - D(x|y): 0.5008 - D(G(z|y)): 0.1550 - KL Divergence: 4.6697
Epoch 22/200
782/782 [==============================] - 116s 148ms/step - d_loss: 0.2856 - g_loss: 2.6794 - D(x|y): 0.5009 - D(G(z|y)): 0.1553 - KL Divergence: 4.7003
Epoch 23/200
782/782 [==============================] - 106s 135ms/step - d_loss: 0.2739 - g_loss: 2.7775 - D(x|y): 0.5005 - D(G(z|y)): 0.1459 - KL Divergence: 4.5943
Epoch 24/200
782/782 [==============================] - 83s 107ms/step - d_loss: 0.2707 - g_loss: 2.8058 - D(x|y): 0.5005 - D(G(z|y)): 0.1473 - KL Divergence: 4.6499
Epoch 25/200
782/782 [==============================] - 96s 122ms/step - d_loss: 0.2928 - g_loss: 2.7117 - D(x|y): 0.4992 - D(G(z|y)): 0.1633 - KL Divergence: 4.6439
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-25.h5
Epoch 26/200
782/782 [==============================] - 96s 122ms/step - d_loss: 0.2652 - g_loss: 2.8549 - D(x|y): 0.5009 - D(G(z|y)): 0.1479 - KL Divergence: 4.8003
Epoch 27/200
782/782 [==============================] - 90s 116ms/step - d_loss: 0.2600 - g_loss: 2.8982 - D(x|y): 0.4998 - D(G(z|y)): 0.1448 - KL Divergence: 4.7303
Epoch 28/200
782/782 [==============================] - 101s 130ms/step - d_loss: 0.2643 - g_loss: 2.8203 - D(x|y): 0.5005 - D(G(z|y)): 0.1520 - KL Divergence: 4.7508
Epoch 29/200
782/782 [==============================] - 98s 125ms/step - d_loss: 0.2583 - g_loss: 2.9410 - D(x|y): 0.5001 - D(G(z|y)): 0.1457 - KL Divergence: 4.7254
Epoch 30/200
782/782 [==============================] - 99s 126ms/step - d_loss: 0.2451 - g_loss: 3.0769 - D(x|y): 0.4996 - D(G(z|y)): 0.1347 - KL Divergence: 4.6875
1/1 [==============================] - 0s 33ms/step

Generator Checkpoint - New ACGAN/generator-epoch-30.h5
Epoch 31/200
782/782 [==============================] - 97s 124ms/step - d_loss: 0.2351 - g_loss: 3.1060 - D(x|y): 0.4995 - D(G(z|y)): 0.1324 - KL Divergence: 4.7472
Epoch 32/200
782/782 [==============================] - 99s 126ms/step - d_loss: 0.2239 - g_loss: 3.3110 - D(x|y): 0.5002 - D(G(z|y)): 0.1253 - KL Divergence: 4.7180
Epoch 33/200
782/782 [==============================] - 108s 138ms/step - d_loss: 0.2178 - g_loss: 3.3819 - D(x|y): 0.5003 - D(G(z|y)): 0.1259 - KL Divergence: 4.7481
Epoch 34/200
782/782 [==============================] - 99s 127ms/step - d_loss: 0.2132 - g_loss: 3.4759 - D(x|y): 0.5011 - D(G(z|y)): 0.1176 - KL Divergence: 4.7260
Epoch 35/200
782/782 [==============================] - 102s 131ms/step - d_loss: 0.2199 - g_loss: 3.4642 - D(x|y): 0.5005 - D(G(z|y)): 0.1236 - KL Divergence: 4.7087
1/1 [==============================] - 0s 30ms/step

Generator Checkpoint - New ACGAN/generator-epoch-35.h5
Epoch 36/200
782/782 [==============================] - 112s 143ms/step - d_loss: 0.2061 - g_loss: 3.5923 - D(x|y): 0.5002 - D(G(z|y)): 0.1141 - KL Divergence: 4.7718
Epoch 37/200
782/782 [==============================] - 95s 122ms/step - d_loss: 0.2149 - g_loss: 3.6089 - D(x|y): 0.5005 - D(G(z|y)): 0.1164 - KL Divergence: 4.7359
Epoch 38/200
782/782 [==============================] - 96s 122ms/step - d_loss: 0.1958 - g_loss: 3.7690 - D(x|y): 0.5003 - D(G(z|y)): 0.1095 - KL Divergence: 4.8173
Epoch 39/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1909 - g_loss: 3.8559 - D(x|y): 0.5003 - D(G(z|y)): 0.1082 - KL Divergence: 4.7473
Epoch 40/200
782/782 [==============================] - 84s 107ms/step - d_loss: 0.1929 - g_loss: 3.9237 - D(x|y): 0.5001 - D(G(z|y)): 0.1063 - KL Divergence: 4.8382
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-40.h5
Epoch 41/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.3209 - g_loss: 3.3620 - D(x|y): 0.5002 - D(G(z|y)): 0.1494 - KL Divergence: 4.7377
Epoch 42/200
782/782 [==============================] - 84s 107ms/step - d_loss: 0.1911 - g_loss: 3.9264 - D(x|y): 0.5009 - D(G(z|y)): 0.1072 - KL Divergence: 4.6980
Epoch 43/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1885 - g_loss: 3.9951 - D(x|y): 0.5007 - D(G(z|y)): 0.1031 - KL Divergence: 4.7631
Epoch 44/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1847 - g_loss: 4.1100 - D(x|y): 0.5004 - D(G(z|y)): 0.1004 - KL Divergence: 4.8429
Epoch 45/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1883 - g_loss: 4.0853 - D(x|y): 0.5006 - D(G(z|y)): 0.1037 - KL Divergence: 4.7726
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-45.h5
Epoch 46/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1819 - g_loss: 4.1744 - D(x|y): 0.4999 - D(G(z|y)): 0.1005 - KL Divergence: 4.7909
Epoch 47/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1773 - g_loss: 4.2348 - D(x|y): 0.5002 - D(G(z|y)): 0.1001 - KL Divergence: 4.7021
Epoch 48/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1705 - g_loss: 4.4048 - D(x|y): 0.5002 - D(G(z|y)): 0.0935 - KL Divergence: 4.7863
Epoch 49/200
782/782 [==============================] - 87s 112ms/step - d_loss: 0.1743 - g_loss: 4.4208 - D(x|y): 0.5007 - D(G(z|y)): 0.0960 - KL Divergence: 4.7784
Epoch 50/200
782/782 [==============================] - 84s 107ms/step - d_loss: 0.1630 - g_loss: 4.4945 - D(x|y): 0.5003 - D(G(z|y)): 0.0915 - KL Divergence: 4.7848
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-50.h5
Epoch 51/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1654 - g_loss: 4.5334 - D(x|y): 0.5009 - D(G(z|y)): 0.0925 - KL Divergence: 4.7669
Epoch 52/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1678 - g_loss: 4.5499 - D(x|y): 0.5003 - D(G(z|y)): 0.0926 - KL Divergence: 4.5998
Epoch 53/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1604 - g_loss: 4.6610 - D(x|y): 0.5010 - D(G(z|y)): 0.0899 - KL Divergence: 4.6990
Epoch 54/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1650 - g_loss: 4.7243 - D(x|y): 0.5000 - D(G(z|y)): 0.0879 - KL Divergence: 4.8778
Epoch 55/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1607 - g_loss: 4.7335 - D(x|y): 0.4996 - D(G(z|y)): 0.0901 - KL Divergence: 4.7149
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-55.h5
Epoch 56/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1627 - g_loss: 4.7537 - D(x|y): 0.5003 - D(G(z|y)): 0.0859 - KL Divergence: 4.6821
Epoch 57/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1559 - g_loss: 4.8275 - D(x|y): 0.5004 - D(G(z|y)): 0.0832 - KL Divergence: 4.7856
Epoch 58/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1535 - g_loss: 4.8816 - D(x|y): 0.5007 - D(G(z|y)): 0.0798 - KL Divergence: 4.7271
Epoch 59/200
782/782 [==============================] - 84s 107ms/step - d_loss: 0.1500 - g_loss: 4.9771 - D(x|y): 0.5009 - D(G(z|y)): 0.0817 - KL Divergence: 4.8199
Epoch 60/200
782/782 [==============================] - 88s 112ms/step - d_loss: 0.1500 - g_loss: 4.9952 - D(x|y): 0.4999 - D(G(z|y)): 0.0815 - KL Divergence: 4.6546
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-60.h5
Epoch 61/200
782/782 [==============================] - 86s 109ms/step - d_loss: 0.1462 - g_loss: 5.0610 - D(x|y): 0.5003 - D(G(z|y)): 0.0808 - KL Divergence: 4.8841
Epoch 62/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1453 - g_loss: 5.2278 - D(x|y): 0.5004 - D(G(z|y)): 0.0763 - KL Divergence: 4.5913
Epoch 63/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1516 - g_loss: 5.1452 - D(x|y): 0.5004 - D(G(z|y)): 0.0792 - KL Divergence: 4.6823
Epoch 64/200
782/782 [==============================] - 89s 114ms/step - d_loss: 0.1469 - g_loss: 5.1944 - D(x|y): 0.5003 - D(G(z|y)): 0.0780 - KL Divergence: 4.6694
Epoch 65/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1439 - g_loss: 5.2272 - D(x|y): 0.5002 - D(G(z|y)): 0.0791 - KL Divergence: 4.7963
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-65.h5
Epoch 66/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1508 - g_loss: 5.2257 - D(x|y): 0.5000 - D(G(z|y)): 0.0777 - KL Divergence: 4.6924
Epoch 67/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1391 - g_loss: 5.3429 - D(x|y): 0.5004 - D(G(z|y)): 0.0753 - KL Divergence: 4.4160
Epoch 68/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1472 - g_loss: 5.3421 - D(x|y): 0.5001 - D(G(z|y)): 0.0777 - KL Divergence: 4.8705
Epoch 69/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1397 - g_loss: 5.4528 - D(x|y): 0.4999 - D(G(z|y)): 0.0721 - KL Divergence: 4.6233
Epoch 70/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1276 - g_loss: 5.5411 - D(x|y): 0.5002 - D(G(z|y)): 0.0698 - KL Divergence: 4.6890
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-70.h5
Epoch 71/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1285 - g_loss: 5.4968 - D(x|y): 0.5003 - D(G(z|y)): 0.0708 - KL Divergence: 4.8176
Epoch 72/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1291 - g_loss: 5.5212 - D(x|y): 0.5001 - D(G(z|y)): 0.0713 - KL Divergence: 4.4357
Epoch 73/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1294 - g_loss: 5.7067 - D(x|y): 0.5003 - D(G(z|y)): 0.0694 - KL Divergence: 4.6374
Epoch 74/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1359 - g_loss: 5.6191 - D(x|y): 0.5007 - D(G(z|y)): 0.0708 - KL Divergence: 4.4713
Epoch 75/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1278 - g_loss: 5.7401 - D(x|y): 0.5002 - D(G(z|y)): 0.0716 - KL Divergence: 4.7517
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-75.h5
Epoch 76/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1378 - g_loss: 5.7418 - D(x|y): 0.5003 - D(G(z|y)): 0.0707 - KL Divergence: 4.6274
Epoch 77/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1273 - g_loss: 5.7199 - D(x|y): 0.5001 - D(G(z|y)): 0.0693 - KL Divergence: 4.6511
Epoch 78/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1281 - g_loss: 5.8291 - D(x|y): 0.4996 - D(G(z|y)): 0.0665 - KL Divergence: 4.5333
Epoch 79/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1251 - g_loss: 5.8501 - D(x|y): 0.5004 - D(G(z|y)): 0.0646 - KL Divergence: 4.5812
Epoch 80/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1225 - g_loss: 5.9108 - D(x|y): 0.4998 - D(G(z|y)): 0.0642 - KL Divergence: 4.7039
1/1 [==============================] - 0s 21ms/step

Generator Checkpoint - New ACGAN/generator-epoch-80.h5
Epoch 81/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1214 - g_loss: 5.9751 - D(x|y): 0.5003 - D(G(z|y)): 0.0649 - KL Divergence: 4.5269
Epoch 82/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.1209 - g_loss: 5.9426 - D(x|y): 0.5001 - D(G(z|y)): 0.0664 - KL Divergence: 4.7336
Epoch 83/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1170 - g_loss: 6.0338 - D(x|y): 0.5002 - D(G(z|y)): 0.0630 - KL Divergence: 4.5350
Epoch 84/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1284 - g_loss: 6.0368 - D(x|y): 0.5007 - D(G(z|y)): 0.0635 - KL Divergence: 4.7853
Epoch 85/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1166 - g_loss: 6.0800 - D(x|y): 0.5006 - D(G(z|y)): 0.0616 - KL Divergence: 4.5917
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-85.h5
Epoch 86/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1221 - g_loss: 6.1005 - D(x|y): 0.5000 - D(G(z|y)): 0.0621 - KL Divergence: 4.7245
Epoch 87/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1205 - g_loss: 6.1454 - D(x|y): 0.5004 - D(G(z|y)): 0.0610 - KL Divergence: 4.6772
Epoch 88/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1162 - g_loss: 6.1645 - D(x|y): 0.5000 - D(G(z|y)): 0.0619 - KL Divergence: 4.5167
Epoch 89/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1151 - g_loss: 6.2019 - D(x|y): 0.5002 - D(G(z|y)): 0.0594 - KL Divergence: 4.4771
Epoch 90/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1173 - g_loss: 6.2635 - D(x|y): 0.5006 - D(G(z|y)): 0.0594 - KL Divergence: 4.6129
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-90.h5
Epoch 91/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1204 - g_loss: 6.1954 - D(x|y): 0.5006 - D(G(z|y)): 0.0617 - KL Divergence: 4.5797
Epoch 92/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1158 - g_loss: 6.2601 - D(x|y): 0.5003 - D(G(z|y)): 0.0572 - KL Divergence: 4.5295
Epoch 93/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1213 - g_loss: 6.2188 - D(x|y): 0.4994 - D(G(z|y)): 0.0598 - KL Divergence: 4.7914
Epoch 94/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1142 - g_loss: 6.2956 - D(x|y): 0.5006 - D(G(z|y)): 0.0604 - KL Divergence: 4.4055
Epoch 95/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1134 - g_loss: 6.3181 - D(x|y): 0.5004 - D(G(z|y)): 0.0596 - KL Divergence: 4.7710
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - New ACGAN/generator-epoch-95.h5
Epoch 96/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1156 - g_loss: 6.3443 - D(x|y): 0.5002 - D(G(z|y)): 0.0573 - KL Divergence: 4.7777
Epoch 97/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1079 - g_loss: 6.4356 - D(x|y): 0.5005 - D(G(z|y)): 0.0582 - KL Divergence: 4.4516
Epoch 98/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1055 - g_loss: 6.4733 - D(x|y): 0.5007 - D(G(z|y)): 0.0562 - KL Divergence: 4.6866
Epoch 99/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1186 - g_loss: 6.4040 - D(x|y): 0.5006 - D(G(z|y)): 0.0592 - KL Divergence: 4.7414
Epoch 100/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1084 - g_loss: 6.5223 - D(x|y): 0.5000 - D(G(z|y)): 0.0544 - KL Divergence: 4.6303
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-100.h5
Epoch 101/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1034 - g_loss: 6.6513 - D(x|y): 0.5006 - D(G(z|y)): 0.0527 - KL Divergence: 4.7078
Epoch 102/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1115 - g_loss: 6.5589 - D(x|y): 0.5005 - D(G(z|y)): 0.0587 - KL Divergence: 4.5346
Epoch 103/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1037 - g_loss: 6.6209 - D(x|y): 0.5002 - D(G(z|y)): 0.0560 - KL Divergence: 4.6566
Epoch 104/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.1182 - g_loss: 6.6285 - D(x|y): 0.5004 - D(G(z|y)): 0.0561 - KL Divergence: 4.5008
Epoch 105/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1034 - g_loss: 6.7122 - D(x|y): 0.5003 - D(G(z|y)): 0.0531 - KL Divergence: 4.7357
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-105.h5
Epoch 106/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.1060 - g_loss: 6.6175 - D(x|y): 0.4998 - D(G(z|y)): 0.0555 - KL Divergence: 4.4286
Epoch 107/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0973 - g_loss: 6.7946 - D(x|y): 0.5000 - D(G(z|y)): 0.0526 - KL Divergence: 4.4410
Epoch 108/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1024 - g_loss: 6.7748 - D(x|y): 0.5001 - D(G(z|y)): 0.0525 - KL Divergence: 4.5522
Epoch 109/200
782/782 [==============================] - 92s 117ms/step - d_loss: 0.1066 - g_loss: 6.8691 - D(x|y): 0.5003 - D(G(z|y)): 0.0522 - KL Divergence: 4.7278
Epoch 110/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.1013 - g_loss: 6.8318 - D(x|y): 0.4997 - D(G(z|y)): 0.0513 - KL Divergence: 4.4754
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-110.h5
Epoch 111/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1022 - g_loss: 6.8294 - D(x|y): 0.5002 - D(G(z|y)): 0.0527 - KL Divergence: 4.6183
Epoch 112/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1055 - g_loss: 6.9106 - D(x|y): 0.5002 - D(G(z|y)): 0.0485 - KL Divergence: 4.6868
Epoch 113/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1006 - g_loss: 6.9977 - D(x|y): 0.4999 - D(G(z|y)): 0.0496 - KL Divergence: 4.4599
Epoch 114/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1004 - g_loss: 6.9846 - D(x|y): 0.5005 - D(G(z|y)): 0.0511 - KL Divergence: 4.7142
Epoch 115/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.1025 - g_loss: 6.8295 - D(x|y): 0.4998 - D(G(z|y)): 0.0539 - KL Divergence: 4.9154
1/1 [==============================] - 0s 33ms/step

Generator Checkpoint - New ACGAN/generator-epoch-115.h5
Epoch 116/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0939 - g_loss: 7.0155 - D(x|y): 0.5000 - D(G(z|y)): 0.0501 - KL Divergence: 4.8058
Epoch 117/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0981 - g_loss: 7.0627 - D(x|y): 0.5004 - D(G(z|y)): 0.0471 - KL Divergence: 4.3865
Epoch 118/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0944 - g_loss: 7.1288 - D(x|y): 0.5001 - D(G(z|y)): 0.0474 - KL Divergence: 4.5317
Epoch 119/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0917 - g_loss: 6.9937 - D(x|y): 0.5007 - D(G(z|y)): 0.0482 - KL Divergence: 4.5877
Epoch 120/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0945 - g_loss: 7.0966 - D(x|y): 0.4999 - D(G(z|y)): 0.0465 - KL Divergence: 4.7501
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-120.h5
Epoch 121/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0921 - g_loss: 7.1253 - D(x|y): 0.5001 - D(G(z|y)): 0.0484 - KL Divergence: 4.5716
Epoch 122/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0987 - g_loss: 7.0774 - D(x|y): 0.5007 - D(G(z|y)): 0.0490 - KL Divergence: 4.6552
Epoch 123/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.0926 - g_loss: 7.1581 - D(x|y): 0.5002 - D(G(z|y)): 0.0472 - KL Divergence: 4.4874
Epoch 124/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0906 - g_loss: 7.1482 - D(x|y): 0.5002 - D(G(z|y)): 0.0474 - KL Divergence: 4.7087
Epoch 125/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0925 - g_loss: 7.1977 - D(x|y): 0.5000 - D(G(z|y)): 0.0470 - KL Divergence: 4.6610
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-125.h5
Epoch 126/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0942 - g_loss: 7.0984 - D(x|y): 0.5004 - D(G(z|y)): 0.0501 - KL Divergence: 4.6217
Epoch 127/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0919 - g_loss: 7.2160 - D(x|y): 0.5000 - D(G(z|y)): 0.0458 - KL Divergence: 4.2903
Epoch 128/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0896 - g_loss: 7.3746 - D(x|y): 0.5003 - D(G(z|y)): 0.0443 - KL Divergence: 4.6586
Epoch 129/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0870 - g_loss: 7.3901 - D(x|y): 0.5005 - D(G(z|y)): 0.0455 - KL Divergence: 4.5553
Epoch 130/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0919 - g_loss: 7.4774 - D(x|y): 0.5000 - D(G(z|y)): 0.0452 - KL Divergence: 4.5623
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-130.h5
Epoch 131/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0871 - g_loss: 7.3757 - D(x|y): 0.5007 - D(G(z|y)): 0.0462 - KL Divergence: 4.5088
Epoch 132/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0879 - g_loss: 7.4005 - D(x|y): 0.4999 - D(G(z|y)): 0.0433 - KL Divergence: 4.5512
Epoch 133/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0818 - g_loss: 7.4716 - D(x|y): 0.5003 - D(G(z|y)): 0.0412 - KL Divergence: 4.3727
Epoch 134/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0855 - g_loss: 7.5332 - D(x|y): 0.5003 - D(G(z|y)): 0.0446 - KL Divergence: 4.7328
Epoch 135/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0894 - g_loss: 7.3797 - D(x|y): 0.5000 - D(G(z|y)): 0.0484 - KL Divergence: 4.6957
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - New ACGAN/generator-epoch-135.h5
Epoch 136/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0920 - g_loss: 7.5068 - D(x|y): 0.5002 - D(G(z|y)): 0.0424 - KL Divergence: 4.8588
Epoch 137/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0914 - g_loss: 7.3360 - D(x|y): 0.5010 - D(G(z|y)): 0.0461 - KL Divergence: 4.6405
Epoch 138/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0832 - g_loss: 7.5442 - D(x|y): 0.5003 - D(G(z|y)): 0.0421 - KL Divergence: 4.6114
Epoch 139/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0891 - g_loss: 7.5231 - D(x|y): 0.5005 - D(G(z|y)): 0.0441 - KL Divergence: 4.5388
Epoch 140/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0817 - g_loss: 7.4989 - D(x|y): 0.5002 - D(G(z|y)): 0.0427 - KL Divergence: 4.5713
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-140.h5
Epoch 141/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0831 - g_loss: 7.7126 - D(x|y): 0.5003 - D(G(z|y)): 0.0399 - KL Divergence: 4.4149
Epoch 142/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0883 - g_loss: 7.5766 - D(x|y): 0.5004 - D(G(z|y)): 0.0424 - KL Divergence: 4.6055
Epoch 143/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0808 - g_loss: 7.5995 - D(x|y): 0.5005 - D(G(z|y)): 0.0410 - KL Divergence: 4.4712
Epoch 144/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0846 - g_loss: 7.7835 - D(x|y): 0.5003 - D(G(z|y)): 0.0389 - KL Divergence: 4.7681
Epoch 145/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.0846 - g_loss: 7.7035 - D(x|y): 0.5004 - D(G(z|y)): 0.0398 - KL Divergence: 4.6552
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-145.h5
Epoch 146/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0783 - g_loss: 7.8423 - D(x|y): 0.5000 - D(G(z|y)): 0.0374 - KL Divergence: 4.5797
Epoch 147/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0869 - g_loss: 7.5578 - D(x|y): 0.5004 - D(G(z|y)): 0.0441 - KL Divergence: 4.4920
Epoch 148/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0825 - g_loss: 7.7799 - D(x|y): 0.5002 - D(G(z|y)): 0.0407 - KL Divergence: 4.5827
Epoch 149/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0828 - g_loss: 7.7144 - D(x|y): 0.5002 - D(G(z|y)): 0.0416 - KL Divergence: 4.4182
Epoch 150/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0826 - g_loss: 7.7520 - D(x|y): 0.5000 - D(G(z|y)): 0.0405 - KL Divergence: 4.6064
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-150.h5
Epoch 151/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0835 - g_loss: 7.6843 - D(x|y): 0.5000 - D(G(z|y)): 0.0419 - KL Divergence: 4.6644
Epoch 152/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0805 - g_loss: 7.7413 - D(x|y): 0.5005 - D(G(z|y)): 0.0380 - KL Divergence: 4.3976
Epoch 153/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0787 - g_loss: 7.7224 - D(x|y): 0.5007 - D(G(z|y)): 0.0397 - KL Divergence: 4.3669
Epoch 154/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0851 - g_loss: 7.8253 - D(x|y): 0.5003 - D(G(z|y)): 0.0385 - KL Divergence: 4.7479
Epoch 155/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0780 - g_loss: 7.6519 - D(x|y): 0.4998 - D(G(z|y)): 0.0388 - KL Divergence: 4.8820
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - New ACGAN/generator-epoch-155.h5
Epoch 156/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0798 - g_loss: 8.0198 - D(x|y): 0.5001 - D(G(z|y)): 0.0374 - KL Divergence: 4.4092
Epoch 157/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0726 - g_loss: 7.8243 - D(x|y): 0.5002 - D(G(z|y)): 0.0382 - KL Divergence: 4.7565
Epoch 158/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0805 - g_loss: 7.8146 - D(x|y): 0.5008 - D(G(z|y)): 0.0396 - KL Divergence: 4.3910
Epoch 159/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0820 - g_loss: 7.7585 - D(x|y): 0.4997 - D(G(z|y)): 0.0387 - KL Divergence: 4.5328
Epoch 160/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0779 - g_loss: 7.9329 - D(x|y): 0.5001 - D(G(z|y)): 0.0362 - KL Divergence: 4.4244
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-160.h5
Epoch 161/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0750 - g_loss: 7.9402 - D(x|y): 0.5004 - D(G(z|y)): 0.0364 - KL Divergence: 4.5819
Epoch 162/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0878 - g_loss: 7.7922 - D(x|y): 0.5001 - D(G(z|y)): 0.0431 - KL Divergence: 4.7076
Epoch 163/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0707 - g_loss: 8.1218 - D(x|y): 0.5003 - D(G(z|y)): 0.0345 - KL Divergence: 4.6556
Epoch 164/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0744 - g_loss: 8.0729 - D(x|y): 0.4995 - D(G(z|y)): 0.0350 - KL Divergence: 4.7184
Epoch 165/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0700 - g_loss: 8.1394 - D(x|y): 0.5002 - D(G(z|y)): 0.0356 - KL Divergence: 4.4522
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New ACGAN/generator-epoch-165.h5
Epoch 166/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0733 - g_loss: 8.0846 - D(x|y): 0.5002 - D(G(z|y)): 0.0344 - KL Divergence: 4.6408
Epoch 167/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0762 - g_loss: 7.9897 - D(x|y): 0.4999 - D(G(z|y)): 0.0364 - KL Divergence: 4.4707
Epoch 168/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0742 - g_loss: 8.0802 - D(x|y): 0.5002 - D(G(z|y)): 0.0366 - KL Divergence: 4.7756
Epoch 169/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0728 - g_loss: 8.1807 - D(x|y): 0.5003 - D(G(z|y)): 0.0352 - KL Divergence: 4.4246
Epoch 170/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0724 - g_loss: 8.2117 - D(x|y): 0.5001 - D(G(z|y)): 0.0345 - KL Divergence: 4.8264
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-170.h5
Epoch 171/200
782/782 [==============================] - 88s 113ms/step - d_loss: 0.0755 - g_loss: 8.0926 - D(x|y): 0.5004 - D(G(z|y)): 0.0378 - KL Divergence: 4.7642
Epoch 172/200
782/782 [==============================] - 87s 112ms/step - d_loss: 0.0702 - g_loss: 8.1292 - D(x|y): 0.5003 - D(G(z|y)): 0.0351 - KL Divergence: 4.3624
Epoch 173/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.0696 - g_loss: 8.1703 - D(x|y): 0.5000 - D(G(z|y)): 0.0334 - KL Divergence: 4.3526
Epoch 174/200
782/782 [==============================] - 87s 112ms/step - d_loss: 0.0684 - g_loss: 8.3260 - D(x|y): 0.5001 - D(G(z|y)): 0.0353 - KL Divergence: 4.8449
Epoch 175/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.0667 - g_loss: 8.0661 - D(x|y): 0.5000 - D(G(z|y)): 0.0348 - KL Divergence: 4.6369
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New ACGAN/generator-epoch-175.h5
Epoch 176/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0651 - g_loss: 8.2996 - D(x|y): 0.5001 - D(G(z|y)): 0.0316 - KL Divergence: 4.8388
Epoch 177/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0716 - g_loss: 8.3022 - D(x|y): 0.5009 - D(G(z|y)): 0.0341 - KL Divergence: 4.5590
Epoch 178/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0660 - g_loss: 8.2865 - D(x|y): 0.5004 - D(G(z|y)): 0.0314 - KL Divergence: 4.6032
Epoch 179/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0684 - g_loss: 8.3598 - D(x|y): 0.5005 - D(G(z|y)): 0.0325 - KL Divergence: 4.5246
Epoch 180/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0710 - g_loss: 8.2936 - D(x|y): 0.5002 - D(G(z|y)): 0.0319 - KL Divergence: 4.5682
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-180.h5
Epoch 181/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0671 - g_loss: 8.3628 - D(x|y): 0.5001 - D(G(z|y)): 0.0321 - KL Divergence: 4.2278
Epoch 182/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0645 - g_loss: 8.5127 - D(x|y): 0.5003 - D(G(z|y)): 0.0295 - KL Divergence: 4.5485
Epoch 183/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0649 - g_loss: 8.6284 - D(x|y): 0.4998 - D(G(z|y)): 0.0297 - KL Divergence: 4.0881
Epoch 184/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0615 - g_loss: 8.4513 - D(x|y): 0.5000 - D(G(z|y)): 0.0286 - KL Divergence: 4.2055
Epoch 185/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0615 - g_loss: 8.6233 - D(x|y): 0.5002 - D(G(z|y)): 0.0287 - KL Divergence: 4.5552
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-185.h5
Epoch 186/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0641 - g_loss: 8.5929 - D(x|y): 0.5001 - D(G(z|y)): 0.0293 - KL Divergence: 4.4563
Epoch 187/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0657 - g_loss: 8.5711 - D(x|y): 0.5002 - D(G(z|y)): 0.0279 - KL Divergence: 4.4411
Epoch 188/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0685 - g_loss: 8.3770 - D(x|y): 0.4999 - D(G(z|y)): 0.0295 - KL Divergence: 4.4341
Epoch 189/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0608 - g_loss: 8.3988 - D(x|y): 0.4999 - D(G(z|y)): 0.0268 - KL Divergence: 4.0901
Epoch 190/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0630 - g_loss: 8.7423 - D(x|y): 0.5000 - D(G(z|y)): 0.0275 - KL Divergence: 4.4028
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New ACGAN/generator-epoch-190.h5
Epoch 191/200
782/782 [==============================] - 86s 110ms/step - d_loss: 0.0656 - g_loss: 8.4579 - D(x|y): 0.5001 - D(G(z|y)): 0.0283 - KL Divergence: 4.5031
Epoch 192/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0614 - g_loss: 8.4956 - D(x|y): 0.5003 - D(G(z|y)): 0.0284 - KL Divergence: 4.6834
Epoch 193/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0619 - g_loss: 8.4405 - D(x|y): 0.5001 - D(G(z|y)): 0.0273 - KL Divergence: 4.4189
Epoch 194/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0604 - g_loss: 8.3121 - D(x|y): 0.5002 - D(G(z|y)): 0.0268 - KL Divergence: 4.7363
Epoch 195/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0695 - g_loss: 8.4642 - D(x|y): 0.5003 - D(G(z|y)): 0.0256 - KL Divergence: 4.8196
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-195.h5
Epoch 196/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0769 - g_loss: 8.3449 - D(x|y): 0.5000 - D(G(z|y)): 0.0278 - KL Divergence: 4.5927
Epoch 197/200
782/782 [==============================] - 85s 109ms/step - d_loss: 0.0763 - g_loss: 7.9450 - D(x|y): 0.5001 - D(G(z|y)): 0.0273 - KL Divergence: 4.2954
Epoch 198/200
782/782 [==============================] - 84s 108ms/step - d_loss: 0.0831 - g_loss: 8.2179 - D(x|y): 0.4999 - D(G(z|y)): 0.0301 - KL Divergence: 4.7499
Epoch 199/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0884 - g_loss: 7.7769 - D(x|y): 0.5007 - D(G(z|y)): 0.0325 - KL Divergence: 4.7332
Epoch 200/200
782/782 [==============================] - 85s 108ms/step - d_loss: 0.0805 - g_loss: 7.7837 - D(x|y): 0.5001 - D(G(z|y)): 0.0294 - KL Divergence: 4.5898
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New ACGAN/generator-epoch-Full Train.h5

Observations

We note that training the model, features like color etc was learnt by the model to generate the images. We also note that the images of some of the animals got blurred and fuzzy at the end. However we can see that unlike with the previous model giving the same image, the NewACGAN did not converge into a state for monotonous output. But we can signs of the output becoming monotonous as the last epochs's trucks look quite similar. Suggesting that using Gaussian Weights and SpectralNormalisations to create the images reduce the changes of model collapsing but made the images generated more fuzzy.

A possible reason why the images are not as good and clear the previous ACGAN images is due to Gaussian weight leading to a slow convergence and poorer images generated. Spectral Normalisation just prevents the modal from collapsing and generate more diverse images.

New ACGAN Evaluation

As mentioned previously we will be using a quantitative and manual evaluation of the ACGAN.

We will start off with a quantitative analysis so that we can find the best model to generate the images for manual evaluation.

# story history object into dataframe
improve_aux_cond_hist_df = pd.DataFrame(improve_aux_cond_hist.history)

# using pandas dataframe to plot out learning curve
fig, (ax1, ax2, ax3) = plt.subplots(
    1, 3, figsize=(24, 8), tight_layout=True)
improve_aux_cond_hist_df.loc[:, ["d_loss", 'g_loss']].plot(
    ax=ax1, title=r'Learning Curve of Loss Function CE')
improve_aux_cond_hist_df.loc[:, ['D(x|y)', 'D(G(z|y))']].plot(
    ax=ax2, title=r'Learning Curve of Discriminator')
improve_aux_cond_hist_df.loc[:, 'KL Divergence'].plot(
    ax=ax3, title=r'Learning Curve of Probabilistic Similarity')
plt.show()

Observations

We see that KL divergence has more spikes in training as the generator steady increases, this is likely due to the Gaussian Weight initialised making training slower. As 180 epochs is the lowest KL divergences, we will be using it as the best generator and show signs of model collapsing.

# Loading Weights for best Generator
saved_weights = 'New ACGAN\generator-epoch-180.h5'
improve_aux_cond_gan.generator.load_weights(saved_weights)
improve_aux_cond_gan.generator.summary()

Model: "generator_ACGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_3 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_3 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_3 (Dense)                (None, 2048)         284672      ['concatenate_3[0][0]']          
                                                                                                  
 batch_normalization_12 (BatchN  (None, 2048)        8192        ['dense_3[0][0]']                
 ormalization)                                                                                    
                                                                                                  
 p_re_lu_4 (PReLU)              (None, 2048)         2048        ['batch_normalization_12[0][0]'] 
                                                                                                  
 reshape_1 (Reshape)            (None, 2, 2, 512)    0           ['p_re_lu_4[0][0]']              
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape_1[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

n = 10000

# generating labels
labels = np.random.randint(low=0, high=10, size=n)
one_hot_labels = to_categorical(labels)

# Generating 1000 Synthetic Images
random_noise = tf.random.normal(shape=(n, NOISE))

synthetic_images = improve_aux_cond_gan.generator.predict(
    [random_noise, one_hot_labels])
print("Latent Vector Dim: {}\tGenerated Images Dim: {}".format(
    random_noise.shape, synthetic_images.shape))

# Scaling back to [0, 1]
synthetic_images -= -1
synthetic_images /= (1 - (-1))

# Display 25 randomly sampled images
fig = plt.figure(figsize=(10, 10), tight_layout=True)
for i in range(25):
    rand_idx = np.random.randint(0, len(synthetic_images))
    ax = fig.add_subplot(5, 5, i+1)
    ax.imshow(synthetic_images[rand_idx])
    ax.set_title(class_labels[labels[rand_idx]])
    ax.axis('off')
plt.show()

313/313 [==============================] - 6s 20ms/step
Latent Vector Dim: (10000, 128)	Generated Images Dim: (10000, 32, 32, 3)

Observations

We see that the shapes of the animals are better but there are certain images where the images are very fuzzy and blurry.

improve_aux_cond_gan_fid_class = GAN_FID(batch_size=512, noise=128,
                                         sample_size=10000, buffer_size=1024, labels=True)
improve_aux_cond_gan_fid_class.fit(
    generator=improve_aux_cond_gan.generator, train_data=x_train)
fid_score = improve_aux_cond_gan_fid_class.evaluate()

Computing Real Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 8s 153ms/step
16/16 [==============================] - 2s 155ms/step
16/16 [==============================] - 2s 156ms/step
16/16 [==============================] - 3s 163ms/step
16/16 [==============================] - 2s 136ms/step
16/16 [==============================] - 2s 136ms/step
16/16 [==============================] - 2s 155ms/step
16/16 [==============================] - 2s 157ms/step
16/16 [==============================] - 2s 154ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 134ms/step
Computing Generated Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 134ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 134ms/step
16/16 [==============================] - 2s 136ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 133ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 138ms/step
16/16 [==============================] - 2s 139ms/step
16/16 [==============================] - 2s 135ms/step
Computed Embeddings	Real Images Embedding Shape: (10240, 2048)	Generated Images Embedding Shape: (10240, 2048)
The computed FID score is: 106.80945871926016

Observations

We see the FID score is higher than the original ACGAN model from 88.7613090961658 to 106.80945871926016. This suggest that the changes made did not improve the model. As mentioned previously, Gaussian weights make the model train more steadily but with a slower speed.

cGAN

New cGAN Discriminator

As mentioned in the DCGAN discriminator creation, we will be using Gaussian Weights for the new ACGAN model.

Gaussian weights will reduce the instability of training [mean = 0, std = 0.02].

# function to create discriminator model
def create_improve_cGAN_discriminator(image_size):
    # input image
    img_input = Input(shape=(image_size), name='Image_Input')
    weights_init = RandomNormal(mean=0, stddev=0.02)

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    base_discriminator = Sequential([
        Conv2D(64, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(128, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(256, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
        Conv2D(512, kernel_size=4, strides=2, padding='same',
               kernel_initializer=weights_init),
        BatchNormalization(momentum=0.8),
        LeakyReLU(0.2),
    ], name='Base_Discriminator')
    discriminator = base_discriminator(img_input)

    discriminator = GlobalAveragePooling2D()(discriminator)

    # Concatenate - combine with conditions y
    merged_layer = Concatenate()([discriminator, conditions])
    discriminator = Dense(512, activation='relu')(merged_layer)

    # Output
    discriminator = Dense(1, activation='sigmoid',
                          name='Output_Layer')(discriminator)

    discriminator = Model(inputs=[img_input, conditions],
                          outputs=discriminator, name='discriminator_cGAN')
    return discriminator


create_improve_cGAN_discriminator(image_size=IMG_SIZE).summary()

Model: "discriminator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Image_Input (InputLayer)       [(None, 32, 32, 3)]  0           []                               
                                                                                                  
 Base_Discriminator (Sequential  (None, 2, 2, 512)   2760384     ['Image_Input[0][0]']            
 )                                                                                                
                                                                                                  
 global_average_pooling2d_4 (Gl  (None, 512)         0           ['Base_Discriminator[0][0]']     
 obalAveragePooling2D)                                                                            
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_7 (Concatenate)    (None, 522)          0           ['global_average_pooling2d_4[0][0
                                                                 ]',                              
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_5 (Dense)                (None, 512)          267776      ['concatenate_7[0][0]']          
                                                                                                  
 Output_Layer (Dense)           (None, 1)            513         ['dense_5[0][0]']                
                                                                                                  
==================================================================================================

c:\Users\Soh Hong Yu\anaconda3\envs\gpu_env\lib\site-packages\keras\initializers\initializers_v2.py:120: UserWarning: The initializer RandomNormal is unseeded and being called multiple times, which will return identical values  each time (even if the initializer is unseeded). Please update your code to provide a seed to the initializer, or avoid using the same initalizer instance more than once.
  warnings.warn(

Total params: 3,028,673
Trainable params: 3,026,753
Non-trainable params: 1,920
__________________________________________________________________________________________________
Trainable params: 3,026,753
Non-trainable params: 1,920
__________________________________________________________________________________________________

plot_model(create_improve_cGAN_discriminator(image_size=IMG_SIZE), to_file="images/models/New cGAN_discriminator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

New cGAN Generator

Similarly, we will use Spectral Normalisation, Gaussian weights and PReLU for the generator. This will allow the generator to adapt better and learn better overall compared to the LeakyReLU used previously.

# function to create generator model
def create_improve_cGAN_generator(noise):

    # gaussian weights initialization
    weights_init = RandomNormal(mean=0, stddev=0.02)

    # latent noise vector z
    z = Input(shape=(noise,), name="Latent_Noise_Vector_z")

    # conditions y
    conditions = Input(shape=(10,), name='Conditions_y')

    # Generator network
    merged_layer = Concatenate()([z, conditions])

    # FC: 2x2x512
    generator = Dense(2*2*512, activation='relu')(merged_layer)
    generator = BatchNormalization(momentum=0.8)(generator)
    generator = PReLU()(generator)
    generator = Reshape((2, 2, 512))(generator)

    base_generator = Sequential([
        # Conv 1: 4x4x256
        SpectralNormalization(Conv2DTranspose(256, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
        # Conv 2: 8x8x128
        SpectralNormalization(Conv2DTranspose(128, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
        # Conv 3: 16x16x64
        SpectralNormalization(Conv2DTranspose(64, kernel_size=4, strides=2,
                                              padding='same', kernel_initializer=weights_init)),
        BatchNormalization(momentum=0.8),
        PReLU(),
    ], name='Base_Generator')
    generator = base_generator(generator)

    # Conv 4: 32x32x3
    generator = Conv2DTranspose(3, kernel_size=4, strides=2, padding='same',
                                activation='tanh', name='Output_Layer')(generator)

    generator = Model(inputs=[z, conditions],
                      outputs=generator,  name='generator_cGAN')
    return generator


create_improve_cGAN_generator(noise=NOISE).summary()

Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_8 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_6 (Dense)                (None, 2048)         284672      ['concatenate_8[0][0]']          
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_8 (Concatenate)    (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_6 (Dense)                (None, 2048)         284672      ['concatenate_8[0][0]']          
                                                                                                  
 batch_normalization_118 (Batch  (None, 2048)        8192        ['dense_6[0][0]']                
 Normalization)                                                                                   
                                                                                                  
 p_re_lu_8 (PReLU)              (None, 2048)         2048        ['batch_normalization_118[0][0]']
                                                                                                  
 reshape_2 (Reshape)            (None, 2, 2, 512)    0           ['p_re_lu_8[0][0]']              
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape_2[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

plot_model(create_improve_cGAN_generator(noise=NOISE), to_file="images/models/New cGAN_generator.png",
           show_shapes=True, show_dtype=True, expand_nested=True, show_layer_activations=True)

New cGAN Callback

class NewcGANMonitor(keras.callbacks.Callback):
    def __init__(self, num_img=20, noise=128, patience=10, vmin=0, vmax=1):
        self.num_img = num_img
        self.noise = noise
        self.patience = patience
        self.vmin = vmin
        self.vmax = vmax
        self.latent_noise_vector = tf.random.normal(
            shape=(self.num_img, self.noise))
        self.conditions = to_categorical([0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                                          0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

    def generate_plot(self):
        # Generate Images
        generated_images = self.model.generator.predict(
            [self.latent_noise_vector, self.conditions])
        # Normalise Image from [vmin, vmax] to [0, 1]
        generated_images -= self.vmin
        generated_images /= (self.vmax - self.vmin)
        row_size = int(np.ceil(self.num_img/5))
        fig = plt.figure(figsize=(10, 2*row_size), tight_layout=True)
        for i in range(self.num_img):
            ax = fig.add_subplot(row_size, 5, i+1)
            ax.imshow(generated_images[i])
            ax.set_title(class_labels[i % 10])
            ax.axis('off')
        plt.show()

    def save_weights(self, epoch=None):
        try:
            if epoch != None:
                name = 'New cGAN/generator-epoch-{}.h5'.format(epoch)
                print('Generator Checkpoint - {}'.format(name))
                self.model.generator.save_weights(
                    filepath=name,
                    save_format='h5'
                )
        except Exception as e:
            print(e)

    def on_epoch_begin(self, epoch, logs=None):
        if epoch % self.patience == 0:
            self.generate_plot()
            self.save_weights(epoch)

    def on_train_end(self, epoch, logs=None):
        self.generate_plot()
        self.save_weights('Full Train')

callbacks = [
    NewcGANMonitor(num_img=20, noise=128, patience=5, vmin=-1, vmax=1),
]

New cGAN Training

class ConditionalGAN(tf.keras.Model):
    def __init__(self, discriminator, generator, noise):
        super(ConditionalGAN, self).__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.noise = noise
        self.gen_loss_tracker = tf.keras.metrics.Mean(name="generator_loss")
        self.disc_loss_tracker = tf.keras.metrics.Mean(
            name="discriminator_loss")
        self.d_xy_tracker = tf.keras.metrics.Mean(name='Mean D(x|y)')
        self.d_g_zy_tracker = tf.keras.metrics.Mean(name='Mean D(G(z|y))')
        self.kl = tf.keras.metrics.KLDivergence()

    def compile(self, d_optimizer, g_optimizer, loss_fn):
        super(ConditionalGAN, self).compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.loss_fn = loss_fn

    def train_step(self, data):
        ### TRAINING DISCRIMINATOR ###
        # Unpack the data.
        real_images, condition = data

        # Sample for latent noise vector z
        batch_size = tf.shape(real_images)[0]
        latent_noise_vector = tf.random.normal(shape=(batch_size, self.noise))

        # Maps the noise latent vector and labels to generate fake images.
        generated_images = self.generator([latent_noise_vector, condition])

        # Combine with real images
        combined_images = tf.concat([generated_images, real_images], axis=0)
        combined_condition = tf.concat([condition, condition], axis=0)

        # Discrimination
        labels = tf.concat(
            [tf.zeros((batch_size, 1)), tf.ones((batch_size, 1))], axis=0
        )

        # Train the discriminator.
        with tf.GradientTape() as tape:
            first_predictions = self.discriminator(
                [combined_images, combined_condition])
            d_loss = self.loss_fn(labels, first_predictions)
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # Computing D(x|y)
        d_xy = tf.math.reduce_mean(first_predictions)

        ### TRAINING GENRATOR ###
        latent_noise_vector = tf.random.normal(shape=(batch_size, self.noise))

        # Assemble labels that say "all real images".
        misleading_labels = tf.ones((batch_size, 1))

        with tf.GradientTape() as tape:
            fake_images = self.generator([latent_noise_vector, condition])
            second_predictions = self.discriminator([fake_images, condition])
            g_loss = self.loss_fn(misleading_labels, second_predictions)
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(
            zip(grads, self.generator.trainable_weights))

        # Computing D(G(z|y))
        d_g_zy = tf.math.reduce_mean(second_predictions)

        # Monitor loss and metrics.
        self.gen_loss_tracker.update_state(g_loss)
        self.disc_loss_tracker.update_state(d_loss)
        self.d_xy_tracker.update_state(d_xy)
        self.d_g_zy_tracker.update_state(d_g_zy)
        self.kl.update_state(real_images, fake_images)

        return {
            "d_loss": self.disc_loss_tracker.result(),
            "g_loss": self.gen_loss_tracker.result(),
            "D(x|y)": self.d_xy_tracker.result(),
            "D(G(z|y))": self.d_g_zy_tracker.result(),
            "KL Divergence": self.
            kl.result(),
        }

tf.keras.backend.clear_session()
K.clear_session()

improve_cond_gan = ConditionalGAN(
    discriminator=create_improve_cGAN_discriminator(image_size=IMG_SIZE),
    generator=create_improve_cGAN_generator(noise=NOISE),
    noise=NOISE
)

improve_cond_gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
    loss_fn=keras.losses.BinaryCrossentropy(),
)

dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
dataset = dataset.shuffle(buffer_size=1024).batch(
    BATCH_SIZE, num_parallel_calls=tf.data.AUTOTUNE).prefetch(tf.data.AUTOTUNE)

improve_cond_gan_hist = improve_cond_gan.fit(
    dataset, epochs=200, use_multiprocessing=True, workers=16, callbacks=callbacks)

1/1 [==============================] - 7s 7s/step

Generator Checkpoint - New cGAN/generator-epoch-0.h5
Epoch 1/200
782/782 [==============================] - 54s 62ms/step - d_loss: 0.2978 - g_loss: 3.3459 - D(x|y): 0.5221 - D(G(z|y)): 0.1088 - KL Divergence: 6.0988
Epoch 2/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.2720 - g_loss: 3.0355 - D(x|y): 0.4933 - D(G(z|y)): 0.0826 - KL Divergence: 6.7751
Epoch 3/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.2494 - g_loss: 3.2739 - D(x|y): 0.4964 - D(G(z|y)): 0.0665 - KL Divergence: 5.6263
Epoch 4/200
782/782 [==============================] - 55s 71ms/step - d_loss: 0.1900 - g_loss: 3.5944 - D(x|y): 0.4976 - D(G(z|y)): 0.0606 - KL Divergence: 6.1849
Epoch 5/200
782/782 [==============================] - 55s 71ms/step - d_loss: 0.2508 - g_loss: 2.9719 - D(x|y): 0.4979 - D(G(z|y)): 0.0801 - KL Divergence: 5.6865
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - New cGAN/generator-epoch-5.h5
Epoch 6/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2900 - g_loss: 2.8784 - D(x|y): 0.5025 - D(G(z|y)): 0.0886 - KL Divergence: 4.3030
Epoch 7/200
782/782 [==============================] - 55s 70ms/step - d_loss: 0.2638 - g_loss: 2.9703 - D(x|y): 0.5016 - D(G(z|y)): 0.0794 - KL Divergence: 4.8253
Epoch 8/200
782/782 [==============================] - 54s 70ms/step - d_loss: 0.3211 - g_loss: 2.6753 - D(x|y): 0.5003 - D(G(z|y)): 0.1039 - KL Divergence: 4.9567
Epoch 9/200
782/782 [==============================] - 53s 68ms/step - d_loss: 0.2964 - g_loss: 2.8504 - D(x|y): 0.5006 - D(G(z|y)): 0.1076 - KL Divergence: 4.7245
Epoch 10/200
782/782 [==============================] - 53s 68ms/step - d_loss: 0.2979 - g_loss: 2.7473 - D(x|y): 0.4999 - D(G(z|y)): 0.1211 - KL Divergence: 4.7051
1/1 [==============================] - 0s 33ms/step

Generator Checkpoint - New cGAN/generator-epoch-10.h5
Epoch 11/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.3139 - g_loss: 2.6440 - D(x|y): 0.4996 - D(G(z|y)): 0.1330 - KL Divergence: 4.6347
Epoch 12/200
782/782 [==============================] - 54s 70ms/step - d_loss: 0.2899 - g_loss: 2.7278 - D(x|y): 0.5001 - D(G(z|y)): 0.1310 - KL Divergence: 4.7942
Epoch 13/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2968 - g_loss: 2.6324 - D(x|y): 0.5011 - D(G(z|y)): 0.1422 - KL Divergence: 4.5704
Epoch 14/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2985 - g_loss: 2.6300 - D(x|y): 0.5006 - D(G(z|y)): 0.1444 - KL Divergence: 4.6995
Epoch 15/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.3298 - g_loss: 2.5145 - D(x|y): 0.4994 - D(G(z|y)): 0.1618 - KL Divergence: 4.6561
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - New cGAN/generator-epoch-15.h5
Epoch 16/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.3063 - g_loss: 2.5317 - D(x|y): 0.5008 - D(G(z|y)): 0.1586 - KL Divergence: 4.6985
Epoch 17/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.3093 - g_loss: 2.5373 - D(x|y): 0.5008 - D(G(z|y)): 0.1588 - KL Divergence: 4.6638
Epoch 18/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2971 - g_loss: 2.6530 - D(x|y): 0.5008 - D(G(z|y)): 0.1519 - KL Divergence: 4.7238
Epoch 19/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2996 - g_loss: 2.5974 - D(x|y): 0.5007 - D(G(z|y)): 0.1553 - KL Divergence: 4.7140
Epoch 20/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2839 - g_loss: 2.6496 - D(x|y): 0.5004 - D(G(z|y)): 0.1508 - KL Divergence: 4.7201
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - New cGAN/generator-epoch-20.h5
Epoch 21/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2788 - g_loss: 2.6916 - D(x|y): 0.5008 - D(G(z|y)): 0.1483 - KL Divergence: 4.8230
Epoch 22/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2773 - g_loss: 2.7279 - D(x|y): 0.5008 - D(G(z|y)): 0.1480 - KL Divergence: 4.7877
Epoch 23/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2898 - g_loss: 2.6814 - D(x|y): 0.5006 - D(G(z|y)): 0.1564 - KL Divergence: 4.7436
Epoch 24/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2827 - g_loss: 2.6113 - D(x|y): 0.5004 - D(G(z|y)): 0.1602 - KL Divergence: 4.7403
Epoch 25/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.2849 - g_loss: 2.6963 - D(x|y): 0.5002 - D(G(z|y)): 0.1573 - KL Divergence: 4.7825
1/1 [==============================] - 0s 30ms/step

Generator Checkpoint - New cGAN/generator-epoch-25.h5
Epoch 26/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.2720 - g_loss: 2.8045 - D(x|y): 0.4997 - D(G(z|y)): 0.1498 - KL Divergence: 4.8034
Epoch 27/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.2653 - g_loss: 2.8807 - D(x|y): 0.5000 - D(G(z|y)): 0.1462 - KL Divergence: 4.6746
Epoch 28/200
782/782 [==============================] - 82s 104ms/step - d_loss: 0.2570 - g_loss: 2.9827 - D(x|y): 0.5004 - D(G(z|y)): 0.1418 - KL Divergence: 4.6800
Epoch 29/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.2506 - g_loss: 3.0585 - D(x|y): 0.4998 - D(G(z|y)): 0.1376 - KL Divergence: 4.7081
Epoch 30/200
782/782 [==============================] - 76s 98ms/step - d_loss: 0.2441 - g_loss: 3.1226 - D(x|y): 0.5007 - D(G(z|y)): 0.1365 - KL Divergence: 4.7073
1/1 [==============================] - 0s 23ms/step

Generator Checkpoint - New cGAN/generator-epoch-30.h5
Epoch 31/200
782/782 [==============================] - 80s 103ms/step - d_loss: 0.2550 - g_loss: 3.0789 - D(x|y): 0.5007 - D(G(z|y)): 0.1429 - KL Divergence: 4.7381
Epoch 32/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.2830 - g_loss: 3.0047 - D(x|y): 0.5008 - D(G(z|y)): 0.1509 - KL Divergence: 4.7244
Epoch 33/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.2281 - g_loss: 3.3576 - D(x|y): 0.5001 - D(G(z|y)): 0.1266 - KL Divergence: 4.7971
Epoch 34/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.2190 - g_loss: 3.4356 - D(x|y): 0.5002 - D(G(z|y)): 0.1221 - KL Divergence: 4.6932
Epoch 35/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.2183 - g_loss: 3.5289 - D(x|y): 0.4996 - D(G(z|y)): 0.1205 - KL Divergence: 4.7403
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New cGAN/generator-epoch-35.h5
Epoch 36/200
782/782 [==============================] - 77s 98ms/step - d_loss: 0.2157 - g_loss: 3.5700 - D(x|y): 0.5006 - D(G(z|y)): 0.1187 - KL Divergence: 4.7593
Epoch 37/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.2169 - g_loss: 3.6181 - D(x|y): 0.5004 - D(G(z|y)): 0.1210 - KL Divergence: 4.6791
Epoch 38/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1977 - g_loss: 3.7868 - D(x|y): 0.5006 - D(G(z|y)): 0.1106 - KL Divergence: 4.7311
Epoch 39/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.1946 - g_loss: 3.8984 - D(x|y): 0.4999 - D(G(z|y)): 0.1069 - KL Divergence: 4.7991
Epoch 40/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1892 - g_loss: 3.9645 - D(x|y): 0.5007 - D(G(z|y)): 0.1052 - KL Divergence: 4.6082
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New cGAN/generator-epoch-40.h5
Epoch 41/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1858 - g_loss: 4.0267 - D(x|y): 0.5007 - D(G(z|y)): 0.1040 - KL Divergence: 4.8212
Epoch 42/200
782/782 [==============================] - 80s 103ms/step - d_loss: 0.1892 - g_loss: 4.0255 - D(x|y): 0.4991 - D(G(z|y)): 0.1025 - KL Divergence: 4.7064
Epoch 43/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.1742 - g_loss: 4.1019 - D(x|y): 0.5000 - D(G(z|y)): 0.1003 - KL Divergence: 4.7073
Epoch 44/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.1789 - g_loss: 4.2403 - D(x|y): 0.5000 - D(G(z|y)): 0.0995 - KL Divergence: 4.6498
Epoch 45/200
782/782 [==============================] - 80s 103ms/step - d_loss: 0.1643 - g_loss: 4.4078 - D(x|y): 0.4999 - D(G(z|y)): 0.0943 - KL Divergence: 4.6208
1/1 [==============================] - 0s 22ms/step

Generator Checkpoint - New cGAN/generator-epoch-45.h5
Epoch 46/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1672 - g_loss: 4.4206 - D(x|y): 0.5001 - D(G(z|y)): 0.0940 - KL Divergence: 4.7090
Epoch 47/200
782/782 [==============================] - 61s 79ms/step - d_loss: 0.1655 - g_loss: 4.4484 - D(x|y): 0.5004 - D(G(z|y)): 0.0921 - KL Divergence: 4.6732
Epoch 48/200
782/782 [==============================] - 57s 73ms/step - d_loss: 0.1604 - g_loss: 4.6099 - D(x|y): 0.5000 - D(G(z|y)): 0.0886 - KL Divergence: 4.5495
Epoch 49/200
782/782 [==============================] - 61s 78ms/step - d_loss: 0.1574 - g_loss: 4.6186 - D(x|y): 0.4998 - D(G(z|y)): 0.0880 - KL Divergence: 4.6346
Epoch 50/200
782/782 [==============================] - 51s 65ms/step - d_loss: 0.1554 - g_loss: 4.6344 - D(x|y): 0.5001 - D(G(z|y)): 0.0880 - KL Divergence: 4.6760
1/1 [==============================] - 0s 30ms/step

Generator Checkpoint - New cGAN/generator-epoch-50.h5
Epoch 51/200
782/782 [==============================] - 59s 75ms/step - d_loss: 0.1605 - g_loss: 4.6824 - D(x|y): 0.5002 - D(G(z|y)): 0.0883 - KL Divergence: 4.4981
Epoch 52/200
782/782 [==============================] - 60s 77ms/step - d_loss: 0.1482 - g_loss: 4.8821 - D(x|y): 0.4997 - D(G(z|y)): 0.0813 - KL Divergence: 4.3949
Epoch 53/200
782/782 [==============================] - 57s 72ms/step - d_loss: 0.1527 - g_loss: 4.8996 - D(x|y): 0.5007 - D(G(z|y)): 0.0854 - KL Divergence: 4.5783
Epoch 54/200
782/782 [==============================] - 54s 70ms/step - d_loss: 0.1501 - g_loss: 4.9697 - D(x|y): 0.5000 - D(G(z|y)): 0.0812 - KL Divergence: 4.4994
Epoch 55/200
782/782 [==============================] - 54s 69ms/step - d_loss: 0.1480 - g_loss: 4.9598 - D(x|y): 0.5003 - D(G(z|y)): 0.0812 - KL Divergence: 4.5852
1/1 [==============================] - 0s 30ms/step

Generator Checkpoint - New cGAN/generator-epoch-55.h5
Epoch 56/200
782/782 [==============================] - 54s 70ms/step - d_loss: 0.1427 - g_loss: 5.0395 - D(x|y): 0.5009 - D(G(z|y)): 0.0771 - KL Divergence: 4.6244
Epoch 57/200
782/782 [==============================] - 55s 70ms/step - d_loss: 0.1407 - g_loss: 5.1246 - D(x|y): 0.5004 - D(G(z|y)): 0.0794 - KL Divergence: 4.6944
Epoch 58/200
782/782 [==============================] - 56s 71ms/step - d_loss: 0.3012 - g_loss: 4.1907 - D(x|y): 0.5017 - D(G(z|y)): 0.1332 - KL Divergence: 4.6172
Epoch 59/200
782/782 [==============================] - 49s 63ms/step - d_loss: 0.1363 - g_loss: 4.9949 - D(x|y): 0.5003 - D(G(z|y)): 0.0756 - KL Divergence: 4.5504
Epoch 60/200
782/782 [==============================] - 55s 71ms/step - d_loss: 0.1384 - g_loss: 5.1037 - D(x|y): 0.5002 - D(G(z|y)): 0.0769 - KL Divergence: 4.6315
1/1 [==============================] - 0s 47ms/step

Generator Checkpoint - New cGAN/generator-epoch-60.h5
Epoch 61/200
782/782 [==============================] - 59s 76ms/step - d_loss: 0.1499 - g_loss: 5.1048 - D(x|y): 0.4997 - D(G(z|y)): 0.0788 - KL Divergence: 4.6915
Epoch 62/200
782/782 [==============================] - 62s 79ms/step - d_loss: 0.1363 - g_loss: 5.2132 - D(x|y): 0.5005 - D(G(z|y)): 0.0758 - KL Divergence: 4.6619
Epoch 63/200
782/782 [==============================] - 61s 78ms/step - d_loss: 0.1366 - g_loss: 5.2235 - D(x|y): 0.5001 - D(G(z|y)): 0.0768 - KL Divergence: 4.7913
Epoch 64/200
782/782 [==============================] - 55s 70ms/step - d_loss: 0.1470 - g_loss: 5.2257 - D(x|y): 0.4997 - D(G(z|y)): 0.0783 - KL Divergence: 4.6809
Epoch 65/200
782/782 [==============================] - 51s 65ms/step - d_loss: 0.1313 - g_loss: 5.3987 - D(x|y): 0.5001 - D(G(z|y)): 0.0731 - KL Divergence: 4.5989
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-65.h5
Epoch 66/200
782/782 [==============================] - 51s 65ms/step - d_loss: 0.1314 - g_loss: 5.4156 - D(x|y): 0.5002 - D(G(z|y)): 0.0730 - KL Divergence: 4.6488
Epoch 67/200
782/782 [==============================] - 55s 70ms/step - d_loss: 0.1276 - g_loss: 5.4792 - D(x|y): 0.5002 - D(G(z|y)): 0.0712 - KL Divergence: 4.5688
Epoch 68/200
782/782 [==============================] - 59s 76ms/step - d_loss: 0.1357 - g_loss: 5.4498 - D(x|y): 0.5007 - D(G(z|y)): 0.0710 - KL Divergence: 4.7601
Epoch 69/200
782/782 [==============================] - 57s 73ms/step - d_loss: 0.1250 - g_loss: 5.5100 - D(x|y): 0.4999 - D(G(z|y)): 0.0698 - KL Divergence: 4.8093
Epoch 70/200
782/782 [==============================] - 57s 73ms/step - d_loss: 0.1275 - g_loss: 5.5687 - D(x|y): 0.5005 - D(G(z|y)): 0.0696 - KL Divergence: 4.5884
1/1 [==============================] - 0s 39ms/step

Generator Checkpoint - New cGAN/generator-epoch-70.h5
Epoch 71/200
782/782 [==============================] - 57s 72ms/step - d_loss: 0.1274 - g_loss: 5.6669 - D(x|y): 0.4999 - D(G(z|y)): 0.0688 - KL Divergence: 4.6138
Epoch 72/200
782/782 [==============================] - 62s 79ms/step - d_loss: 0.1258 - g_loss: 5.7240 - D(x|y): 0.5004 - D(G(z|y)): 0.0683 - KL Divergence: 4.6379
Epoch 73/200
782/782 [==============================] - 59s 76ms/step - d_loss: 0.1209 - g_loss: 5.8340 - D(x|y): 0.5000 - D(G(z|y)): 0.0657 - KL Divergence: 4.6185
Epoch 74/200
782/782 [==============================] - 58s 74ms/step - d_loss: 0.1214 - g_loss: 5.8127 - D(x|y): 0.5004 - D(G(z|y)): 0.0649 - KL Divergence: 4.5770
Epoch 75/200
782/782 [==============================] - 64s 82ms/step - d_loss: 0.1245 - g_loss: 5.8535 - D(x|y): 0.5005 - D(G(z|y)): 0.0651 - KL Divergence: 4.7295
1/1 [==============================] - 0s 77ms/step

Generator Checkpoint - New cGAN/generator-epoch-75.h5
Epoch 76/200
782/782 [==============================] - 59s 76ms/step - d_loss: 0.1164 - g_loss: 5.9926 - D(x|y): 0.5002 - D(G(z|y)): 0.0653 - KL Divergence: 4.6654
Epoch 77/200
782/782 [==============================] - 58s 75ms/step - d_loss: 0.1216 - g_loss: 5.9395 - D(x|y): 0.4997 - D(G(z|y)): 0.0665 - KL Divergence: 4.6298
Epoch 78/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1142 - g_loss: 6.0926 - D(x|y): 0.5001 - D(G(z|y)): 0.0598 - KL Divergence: 4.6145
Epoch 79/200
782/782 [==============================] - 56s 71ms/step - d_loss: 0.1221 - g_loss: 5.9829 - D(x|y): 0.5003 - D(G(z|y)): 0.0650 - KL Divergence: 4.6439
Epoch 80/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1159 - g_loss: 6.0471 - D(x|y): 0.5003 - D(G(z|y)): 0.0630 - KL Divergence: 4.6721
1/1 [==============================] - 0s 37ms/step

Generator Checkpoint - New cGAN/generator-epoch-80.h5
Epoch 81/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1146 - g_loss: 6.1334 - D(x|y): 0.5003 - D(G(z|y)): 0.0597 - KL Divergence: 4.4338
Epoch 82/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1171 - g_loss: 6.0960 - D(x|y): 0.4996 - D(G(z|y)): 0.0635 - KL Divergence: 4.6909
Epoch 83/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1115 - g_loss: 6.2260 - D(x|y): 0.5000 - D(G(z|y)): 0.0613 - KL Divergence: 4.6485
Epoch 84/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1138 - g_loss: 6.2236 - D(x|y): 0.5003 - D(G(z|y)): 0.0604 - KL Divergence: 4.5212
Epoch 85/200
782/782 [==============================] - 56s 72ms/step - d_loss: 0.1212 - g_loss: 6.1302 - D(x|y): 0.4998 - D(G(z|y)): 0.0597 - KL Divergence: 4.5620
1/1 [==============================] - 0s 37ms/step

Generator Checkpoint - New cGAN/generator-epoch-85.h5
Epoch 86/200
782/782 [==============================] - 58s 74ms/step - d_loss: 0.1112 - g_loss: 6.2619 - D(x|y): 0.5004 - D(G(z|y)): 0.0577 - KL Divergence: 4.7000
Epoch 87/200
782/782 [==============================] - 65s 83ms/step - d_loss: 0.1115 - g_loss: 6.3115 - D(x|y): 0.5002 - D(G(z|y)): 0.0604 - KL Divergence: 4.7278
Epoch 88/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.1073 - g_loss: 6.3657 - D(x|y): 0.5003 - D(G(z|y)): 0.0575 - KL Divergence: 4.6410
Epoch 89/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.1094 - g_loss: 6.3513 - D(x|y): 0.5008 - D(G(z|y)): 0.0578 - KL Divergence: 4.6047
Epoch 90/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.1095 - g_loss: 6.3270 - D(x|y): 0.5002 - D(G(z|y)): 0.0589 - KL Divergence: 4.5860
1/1 [==============================] - 0s 51ms/step

Generator Checkpoint - New cGAN/generator-epoch-90.h5
Epoch 91/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1058 - g_loss: 6.4651 - D(x|y): 0.4999 - D(G(z|y)): 0.0575 - KL Divergence: 4.5938
Epoch 92/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1043 - g_loss: 6.4436 - D(x|y): 0.5003 - D(G(z|y)): 0.0581 - KL Divergence: 4.5124
Epoch 93/200
782/782 [==============================] - 80s 102ms/step - d_loss: 0.1143 - g_loss: 6.3768 - D(x|y): 0.5003 - D(G(z|y)): 0.0608 - KL Divergence: 4.5867
Epoch 94/200
782/782 [==============================] - 80s 103ms/step - d_loss: 0.1055 - g_loss: 6.4705 - D(x|y): 0.5005 - D(G(z|y)): 0.0570 - KL Divergence: 4.6591
Epoch 95/200
782/782 [==============================] - 83s 106ms/step - d_loss: 0.1087 - g_loss: 6.4679 - D(x|y): 0.5003 - D(G(z|y)): 0.0570 - KL Divergence: 4.5465
1/1 [==============================] - 0s 82ms/step

Generator Checkpoint - New cGAN/generator-epoch-95.h5
Epoch 96/200
782/782 [==============================] - 87s 111ms/step - d_loss: 0.1072 - g_loss: 6.5301 - D(x|y): 0.5007 - D(G(z|y)): 0.0565 - KL Divergence: 4.6018
Epoch 97/200
782/782 [==============================] - 90s 116ms/step - d_loss: 0.1100 - g_loss: 6.6123 - D(x|y): 0.5001 - D(G(z|y)): 0.0557 - KL Divergence: 4.6057
Epoch 98/200
782/782 [==============================] - 86s 109ms/step - d_loss: 0.1081 - g_loss: 6.5173 - D(x|y): 0.5004 - D(G(z|y)): 0.0556 - KL Divergence: 4.7520
Epoch 99/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.1057 - g_loss: 6.5616 - D(x|y): 0.5002 - D(G(z|y)): 0.0549 - KL Divergence: 4.6622
Epoch 100/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.1048 - g_loss: 6.6062 - D(x|y): 0.4998 - D(G(z|y)): 0.0541 - KL Divergence: 4.6291
1/1 [==============================] - 0s 59ms/step

Generator Checkpoint - New cGAN/generator-epoch-100.h5
Epoch 101/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.1006 - g_loss: 6.6132 - D(x|y): 0.5010 - D(G(z|y)): 0.0528 - KL Divergence: 4.6096
Epoch 102/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.1058 - g_loss: 6.6935 - D(x|y): 0.5000 - D(G(z|y)): 0.0540 - KL Divergence: 4.7265
Epoch 103/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.0948 - g_loss: 6.7738 - D(x|y): 0.5002 - D(G(z|y)): 0.0523 - KL Divergence: 4.7126
Epoch 104/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.1006 - g_loss: 6.7151 - D(x|y): 0.5009 - D(G(z|y)): 0.0521 - KL Divergence: 4.5973
Epoch 105/200
782/782 [==============================] - 82s 104ms/step - d_loss: 0.1090 - g_loss: 6.7528 - D(x|y): 0.5000 - D(G(z|y)): 0.0547 - KL Divergence: 4.6664
1/1 [==============================] - 0s 59ms/step

Generator Checkpoint - New cGAN/generator-epoch-105.h5
Epoch 106/200
782/782 [==============================] - 81s 104ms/step - d_loss: 0.1080 - g_loss: 6.6742 - D(x|y): 0.5003 - D(G(z|y)): 0.0553 - KL Divergence: 4.5811
Epoch 107/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.1012 - g_loss: 6.8466 - D(x|y): 0.5003 - D(G(z|y)): 0.0510 - KL Divergence: 4.7239
Epoch 108/200
782/782 [==============================] - 82s 104ms/step - d_loss: 0.0990 - g_loss: 6.7692 - D(x|y): 0.5000 - D(G(z|y)): 0.0499 - KL Divergence: 4.5766
Epoch 109/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.1039 - g_loss: 6.8116 - D(x|y): 0.5000 - D(G(z|y)): 0.0506 - KL Divergence: 4.5668
Epoch 110/200
782/782 [==============================] - 81s 103ms/step - d_loss: 0.0994 - g_loss: 6.8772 - D(x|y): 0.5003 - D(G(z|y)): 0.0514 - KL Divergence: 4.6859
1/1 [==============================] - 0s 49ms/step

Generator Checkpoint - New cGAN/generator-epoch-110.h5
Epoch 111/200
782/782 [==============================] - 82s 105ms/step - d_loss: 0.0944 - g_loss: 6.9157 - D(x|y): 0.5002 - D(G(z|y)): 0.0489 - KL Divergence: 4.6663
Epoch 112/200
782/782 [==============================] - 66s 84ms/step - d_loss: 0.0959 - g_loss: 6.9419 - D(x|y): 0.4999 - D(G(z|y)): 0.0495 - KL Divergence: 4.6097
Epoch 113/200
782/782 [==============================] - 45s 58ms/step - d_loss: 0.0985 - g_loss: 7.0470 - D(x|y): 0.5004 - D(G(z|y)): 0.0492 - KL Divergence: 4.4436
Epoch 114/200
782/782 [==============================] - 45s 58ms/step - d_loss: 0.0974 - g_loss: 7.0387 - D(x|y): 0.5002 - D(G(z|y)): 0.0494 - KL Divergence: 4.5173
Epoch 115/200
782/782 [==============================] - 46s 58ms/step - d_loss: 0.0990 - g_loss: 6.9390 - D(x|y): 0.4999 - D(G(z|y)): 0.0510 - KL Divergence: 4.5969
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-115.h5
Epoch 116/200
782/782 [==============================] - 45s 57ms/step - d_loss: 0.0993 - g_loss: 7.0117 - D(x|y): 0.5012 - D(G(z|y)): 0.0513 - KL Divergence: 4.6055
Epoch 117/200
782/782 [==============================] - 45s 57ms/step - d_loss: 0.0929 - g_loss: 7.1218 - D(x|y): 0.5001 - D(G(z|y)): 0.0481 - KL Divergence: 4.6520
Epoch 118/200
782/782 [==============================] - 45s 58ms/step - d_loss: 0.0956 - g_loss: 7.0729 - D(x|y): 0.5003 - D(G(z|y)): 0.0496 - KL Divergence: 4.6151
Epoch 119/200
782/782 [==============================] - 51s 65ms/step - d_loss: 0.0941 - g_loss: 7.0706 - D(x|y): 0.5001 - D(G(z|y)): 0.0495 - KL Divergence: 4.5493
Epoch 120/200
782/782 [==============================] - 51s 66ms/step - d_loss: 0.0879 - g_loss: 7.2113 - D(x|y): 0.4998 - D(G(z|y)): 0.0457 - KL Divergence: 4.5559
1/1 [==============================] - 0s 34ms/step

Generator Checkpoint - New cGAN/generator-epoch-120.h5
Epoch 121/200
782/782 [==============================] - 52s 67ms/step - d_loss: 0.0961 - g_loss: 7.1849 - D(x|y): 0.4999 - D(G(z|y)): 0.0485 - KL Divergence: 4.6121
Epoch 122/200
782/782 [==============================] - 49s 63ms/step - d_loss: 0.0906 - g_loss: 7.2247 - D(x|y): 0.5002 - D(G(z|y)): 0.0481 - KL Divergence: 4.5534
Epoch 123/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0936 - g_loss: 7.2385 - D(x|y): 0.5000 - D(G(z|y)): 0.0454 - KL Divergence: 4.6024
Epoch 124/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0815 - g_loss: 7.3134 - D(x|y): 0.5003 - D(G(z|y)): 0.0442 - KL Divergence: 4.6128
Epoch 125/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0886 - g_loss: 7.3857 - D(x|y): 0.5001 - D(G(z|y)): 0.0458 - KL Divergence: 4.7411
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - New cGAN/generator-epoch-125.h5
Epoch 126/200
782/782 [==============================] - 49s 62ms/step - d_loss: 0.0898 - g_loss: 7.2204 - D(x|y): 0.5003 - D(G(z|y)): 0.0466 - KL Divergence: 4.3763
Epoch 127/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0904 - g_loss: 7.2214 - D(x|y): 0.5001 - D(G(z|y)): 0.0476 - KL Divergence: 4.5590
Epoch 128/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0883 - g_loss: 7.3766 - D(x|y): 0.5006 - D(G(z|y)): 0.0451 - KL Divergence: 4.5205
Epoch 129/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0865 - g_loss: 7.4122 - D(x|y): 0.4998 - D(G(z|y)): 0.0446 - KL Divergence: 4.4949
Epoch 130/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0905 - g_loss: 7.3925 - D(x|y): 0.5010 - D(G(z|y)): 0.0476 - KL Divergence: 4.5745
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - New cGAN/generator-epoch-130.h5
Epoch 131/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0924 - g_loss: 7.3476 - D(x|y): 0.5002 - D(G(z|y)): 0.0461 - KL Divergence: 4.5651
Epoch 132/200
782/782 [==============================] - 48s 61ms/step - d_loss: 0.0846 - g_loss: 7.5804 - D(x|y): 0.5000 - D(G(z|y)): 0.0416 - KL Divergence: 4.6822
Epoch 133/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0896 - g_loss: 7.5211 - D(x|y): 0.4998 - D(G(z|y)): 0.0449 - KL Divergence: 4.6315
Epoch 134/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0882 - g_loss: 7.4527 - D(x|y): 0.5001 - D(G(z|y)): 0.0465 - KL Divergence: 4.5537
Epoch 135/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0880 - g_loss: 7.5451 - D(x|y): 0.5003 - D(G(z|y)): 0.0441 - KL Divergence: 4.6983
1/1 [==============================] - 0s 29ms/step

Generator Checkpoint - New cGAN/generator-epoch-135.h5
Epoch 136/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0824 - g_loss: 7.6167 - D(x|y): 0.5004 - D(G(z|y)): 0.0428 - KL Divergence: 4.6904
Epoch 137/200
782/782 [==============================] - 49s 62ms/step - d_loss: 0.0813 - g_loss: 7.6924 - D(x|y): 0.5009 - D(G(z|y)): 0.0428 - KL Divergence: 4.7189
Epoch 138/200
782/782 [==============================] - 49s 62ms/step - d_loss: 0.0868 - g_loss: 7.6051 - D(x|y): 0.5002 - D(G(z|y)): 0.0422 - KL Divergence: 4.6432
Epoch 139/200
782/782 [==============================] - 49s 62ms/step - d_loss: 0.0889 - g_loss: 7.5492 - D(x|y): 0.5002 - D(G(z|y)): 0.0466 - KL Divergence: 4.4963
Epoch 140/200
782/782 [==============================] - 49s 62ms/step - d_loss: 0.0835 - g_loss: 7.5683 - D(x|y): 0.5001 - D(G(z|y)): 0.0416 - KL Divergence: 4.5735
1/1 [==============================] - 0s 31ms/step

Generator Checkpoint - New cGAN/generator-epoch-140.h5
Epoch 141/200
782/782 [==============================] - 49s 63ms/step - d_loss: 0.0883 - g_loss: 7.5399 - D(x|y): 0.5003 - D(G(z|y)): 0.0443 - KL Divergence: 4.4941
Epoch 142/200
782/782 [==============================] - 48s 62ms/step - d_loss: 0.0847 - g_loss: 7.7248 - D(x|y): 0.5000 - D(G(z|y)): 0.0422 - KL Divergence: 4.5994
Epoch 143/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0837 - g_loss: 7.6992 - D(x|y): 0.5000 - D(G(z|y)): 0.0402 - KL Divergence: 4.6784
Epoch 144/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0853 - g_loss: 7.6384 - D(x|y): 0.5004 - D(G(z|y)): 0.0440 - KL Divergence: 4.6025
Epoch 145/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0865 - g_loss: 7.7462 - D(x|y): 0.5003 - D(G(z|y)): 0.0417 - KL Divergence: 4.6686
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-145.h5
Epoch 146/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0772 - g_loss: 7.9372 - D(x|y): 0.4997 - D(G(z|y)): 0.0376 - KL Divergence: 4.7923
Epoch 147/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0808 - g_loss: 7.8268 - D(x|y): 0.5000 - D(G(z|y)): 0.0412 - KL Divergence: 4.6729
Epoch 148/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0890 - g_loss: 7.6309 - D(x|y): 0.5005 - D(G(z|y)): 0.0432 - KL Divergence: 4.5258
Epoch 149/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0834 - g_loss: 7.6751 - D(x|y): 0.5005 - D(G(z|y)): 0.0414 - KL Divergence: 4.6080
Epoch 150/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0815 - g_loss: 7.8251 - D(x|y): 0.5002 - D(G(z|y)): 0.0390 - KL Divergence: 4.8169
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-150.h5
Epoch 151/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0843 - g_loss: 7.8169 - D(x|y): 0.5002 - D(G(z|y)): 0.0416 - KL Divergence: 4.7616
Epoch 152/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0826 - g_loss: 7.7697 - D(x|y): 0.5003 - D(G(z|y)): 0.0405 - KL Divergence: 4.5633
Epoch 153/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0780 - g_loss: 7.8280 - D(x|y): 0.5000 - D(G(z|y)): 0.0376 - KL Divergence: 4.5700
Epoch 154/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0858 - g_loss: 7.8240 - D(x|y): 0.5001 - D(G(z|y)): 0.0410 - KL Divergence: 4.5518
Epoch 155/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0767 - g_loss: 7.8748 - D(x|y): 0.5003 - D(G(z|y)): 0.0375 - KL Divergence: 4.5938
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New cGAN/generator-epoch-155.h5
Epoch 156/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0816 - g_loss: 7.9283 - D(x|y): 0.5005 - D(G(z|y)): 0.0385 - KL Divergence: 4.5177
Epoch 157/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0793 - g_loss: 8.0286 - D(x|y): 0.5005 - D(G(z|y)): 0.0378 - KL Divergence: 4.5369
Epoch 158/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0815 - g_loss: 7.8625 - D(x|y): 0.5001 - D(G(z|y)): 0.0405 - KL Divergence: 4.7889
Epoch 159/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0794 - g_loss: 7.8124 - D(x|y): 0.4999 - D(G(z|y)): 0.0415 - KL Divergence: 4.5146
Epoch 160/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0757 - g_loss: 7.8585 - D(x|y): 0.4996 - D(G(z|y)): 0.0385 - KL Divergence: 4.2978
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-160.h5
Epoch 161/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0837 - g_loss: 7.9510 - D(x|y): 0.5000 - D(G(z|y)): 0.0387 - KL Divergence: 4.3841
Epoch 162/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0821 - g_loss: 8.1019 - D(x|y): 0.5002 - D(G(z|y)): 0.0391 - KL Divergence: 4.3928
Epoch 163/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0754 - g_loss: 8.1943 - D(x|y): 0.5000 - D(G(z|y)): 0.0361 - KL Divergence: 4.6872
Epoch 164/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0689 - g_loss: 8.2363 - D(x|y): 0.5003 - D(G(z|y)): 0.0346 - KL Divergence: 4.7921
Epoch 165/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0709 - g_loss: 8.1513 - D(x|y): 0.5005 - D(G(z|y)): 0.0362 - KL Divergence: 4.6852
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New cGAN/generator-epoch-165.h5
Epoch 166/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0737 - g_loss: 8.0642 - D(x|y): 0.5005 - D(G(z|y)): 0.0379 - KL Divergence: 4.4787
Epoch 167/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0745 - g_loss: 8.0499 - D(x|y): 0.5003 - D(G(z|y)): 0.0389 - KL Divergence: 4.7231
Epoch 168/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0765 - g_loss: 7.9880 - D(x|y): 0.5003 - D(G(z|y)): 0.0364 - KL Divergence: 4.3611
Epoch 169/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0775 - g_loss: 8.1135 - D(x|y): 0.5001 - D(G(z|y)): 0.0345 - KL Divergence: 4.2059
Epoch 170/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0764 - g_loss: 8.1005 - D(x|y): 0.4999 - D(G(z|y)): 0.0377 - KL Divergence: 4.4249
1/1 [==============================] - 0s 27ms/step

Generator Checkpoint - New cGAN/generator-epoch-170.h5
Epoch 171/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0806 - g_loss: 8.3493 - D(x|y): 0.4998 - D(G(z|y)): 0.0366 - KL Divergence: 4.7445
Epoch 172/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0732 - g_loss: 8.1573 - D(x|y): 0.5001 - D(G(z|y)): 0.0365 - KL Divergence: 4.5028
Epoch 173/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0692 - g_loss: 8.3394 - D(x|y): 0.5004 - D(G(z|y)): 0.0344 - KL Divergence: 4.4728
Epoch 174/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0700 - g_loss: 8.3384 - D(x|y): 0.5001 - D(G(z|y)): 0.0348 - KL Divergence: 4.3912
Epoch 175/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0719 - g_loss: 8.2288 - D(x|y): 0.4997 - D(G(z|y)): 0.0330 - KL Divergence: 4.3775
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-175.h5
Epoch 176/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0674 - g_loss: 8.3023 - D(x|y): 0.4999 - D(G(z|y)): 0.0342 - KL Divergence: 4.5305
Epoch 177/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0769 - g_loss: 8.0972 - D(x|y): 0.4992 - D(G(z|y)): 0.0367 - KL Divergence: 4.5238
Epoch 178/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0632 - g_loss: 8.4507 - D(x|y): 0.5000 - D(G(z|y)): 0.0325 - KL Divergence: 4.6332
Epoch 179/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0692 - g_loss: 8.1587 - D(x|y): 0.4998 - D(G(z|y)): 0.0344 - KL Divergence: 4.7506
Epoch 180/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0675 - g_loss: 8.3777 - D(x|y): 0.4996 - D(G(z|y)): 0.0301 - KL Divergence: 4.4855
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-180.h5
Epoch 181/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0678 - g_loss: 8.5439 - D(x|y): 0.5003 - D(G(z|y)): 0.0307 - KL Divergence: 4.7544
Epoch 182/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0722 - g_loss: 8.8189 - D(x|y): 0.5002 - D(G(z|y)): 0.0297 - KL Divergence: 5.1058
Epoch 183/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0682 - g_loss: 8.4843 - D(x|y): 0.5005 - D(G(z|y)): 0.0279 - KL Divergence: 4.8050
Epoch 184/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0744 - g_loss: 8.5693 - D(x|y): 0.5000 - D(G(z|y)): 0.0318 - KL Divergence: 4.3674
Epoch 185/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0748 - g_loss: 8.4157 - D(x|y): 0.5001 - D(G(z|y)): 0.0320 - KL Divergence: 4.7888
1/1 [==============================] - 0s 28ms/step

Generator Checkpoint - New cGAN/generator-epoch-185.h5
Epoch 186/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0674 - g_loss: 8.5851 - D(x|y): 0.4999 - D(G(z|y)): 0.0307 - KL Divergence: 4.8173
Epoch 187/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0689 - g_loss: 8.6554 - D(x|y): 0.5006 - D(G(z|y)): 0.0301 - KL Divergence: 4.7703
Epoch 188/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0638 - g_loss: 8.2742 - D(x|y): 0.5001 - D(G(z|y)): 0.0279 - KL Divergence: 4.6779
Epoch 189/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0690 - g_loss: 8.2845 - D(x|y): 0.5004 - D(G(z|y)): 0.0300 - KL Divergence: 4.8092
Epoch 190/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0676 - g_loss: 8.6302 - D(x|y): 0.5004 - D(G(z|y)): 0.0304 - KL Divergence: 4.5259
1/1 [==============================] - 0s 26ms/step

Generator Checkpoint - New cGAN/generator-epoch-190.h5
Epoch 191/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0712 - g_loss: 8.2033 - D(x|y): 0.4999 - D(G(z|y)): 0.0330 - KL Divergence: 4.4581
Epoch 192/200
782/782 [==============================] - 47s 60ms/step - d_loss: 0.0613 - g_loss: 8.3229 - D(x|y): 0.5003 - D(G(z|y)): 0.0280 - KL Divergence: 4.6061
Epoch 193/200
782/782 [==============================] - 46s 59ms/step - d_loss: 0.0628 - g_loss: 8.5700 - D(x|y): 0.5003 - D(G(z|y)): 0.0294 - KL Divergence: 4.5541
Epoch 194/200
782/782 [==============================] - 45s 58ms/step - d_loss: 0.0654 - g_loss: 8.2692 - D(x|y): 0.5003 - D(G(z|y)): 0.0296 - KL Divergence: 4.7508
Epoch 195/200
782/782 [==============================] - 44s 56ms/step - d_loss: 0.0648 - g_loss: 8.3486 - D(x|y): 0.5002 - D(G(z|y)): 0.0285 - KL Divergence: 4.2113
1/1 [==============================] - 0s 24ms/step

Generator Checkpoint - New cGAN/generator-epoch-195.h5
Epoch 196/200
782/782 [==============================] - 44s 57ms/step - d_loss: 0.0693 - g_loss: 8.6436 - D(x|y): 0.5001 - D(G(z|y)): 0.0313 - KL Divergence: 4.1144
Epoch 197/200
782/782 [==============================] - 44s 56ms/step - d_loss: 0.0653 - g_loss: 8.1464 - D(x|y): 0.5001 - D(G(z|y)): 0.0312 - KL Divergence: 4.6433
Epoch 198/200
782/782 [==============================] - 45s 57ms/step - d_loss: 0.0634 - g_loss: 8.2508 - D(x|y): 0.5000 - D(G(z|y)): 0.0292 - KL Divergence: 4.7606
Epoch 199/200
782/782 [==============================] - 44s 57ms/step - d_loss: 0.0652 - g_loss: 8.3266 - D(x|y): 0.5002 - D(G(z|y)): 0.0283 - KL Divergence: 4.7572
Epoch 200/200
782/782 [==============================] - 44s 57ms/step - d_loss: 0.0699 - g_loss: 8.2233 - D(x|y): 0.5005 - D(G(z|y)): 0.0306 - KL Divergence: 4.0840
1/1 [==============================] - 0s 25ms/step

Generator Checkpoint - New cGAN/generator-epoch-Full Train.h5

Observations

At around 180 - 185 epochs, we note that the models have collapsed as Truck and Automobile starts repeating the image that was generated. Around 190 epochs, the generator starts generating images of Truck and Automobile that is exactly the same. We note that other class looks generated looks good.

New cGAN Evaluation

As mentioned previously we will be using a quantitative and manual evaluation of the cGAN.

We will start off with a quantitative analysis so that we can find the best model to generate the images for manual evaluation.

# story history object into dataframe
improve_cond_gan_hist_df = pd.DataFrame(improve_cond_gan_hist.history)

# using pandas dataframe to plot out learning curve
fig, (ax1, ax2, ax3) = plt.subplots(
    1, 3, figsize=(24, 8), tight_layout=True)
improve_cond_gan_hist_df.loc[:, ["d_loss", 'g_loss']].plot(
    ax=ax1, title=r'Learning Curve of Loss Function CE')
improve_cond_gan_hist_df.loc[:, ['D(x|y)', 'D(G(z|y))']].plot(
    ax=ax2, title=r'Learning Curve of Discriminator')
improve_cond_gan_hist_df.loc[:, 'KL Divergence'].plot(
    ax=ax3, title=r'Learning Curve of Probabilistic Similarity')
plt.show()

Observations

We see that KL divergence has more spikes in training as the generator steady decrease, this is likely due to the Gaussian Weight initialised making training slower. As 180 epochs the model collapsed, the lowest KL Divergence score is at 160 epochs.

# Loading Weights for best Generator
saved_weights = 'New cGAN\generator-epoch-160.h5'
improve_cond_gan.generator.load_weights(saved_weights)
improve_cond_gan.generator.summary()

Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_10 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_8 (Dense)                (None, 2048)         284672      ['concatenate_10[0][0]']         
                                                                                                  
 batch_normalization_126 (Batch  (None, 2048)        8192        ['dense_8[0][0]']                
 Normalization)                                                                                   
                                                                                                  
 p_re_lu_12 (PReLU)             (None, 2048)         2048        ['batch_normalization_126[0][0]']
                                                                                                  
 reshape_3 (Reshape)            (None, 2, 2, 512)    0           ['p_re_lu_12[0][0]']             
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape_3[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

n = 10000

# generating labels
labels = np.random.randint(low=0, high=10, size=n)
one_hot_labels = to_categorical(labels)

# Generating 1000 Synthetic Images
random_noise = tf.random.normal(shape=(n, NOISE))

synthetic_images = improve_cond_gan.generator.predict(
    [random_noise, one_hot_labels])
print("Latent Vector Dim: {}\tGenerated Images Dim: {}".format(
    random_noise.shape, synthetic_images.shape))

# Scaling back to [0, 1]
synthetic_images -= -1
synthetic_images /= (1 - (-1))

# Display 25 randomly sampled images
fig = plt.figure(figsize=(10, 10), tight_layout=True)
for i in range(25):
    rand_idx = np.random.randint(0, len(synthetic_images))
    ax = fig.add_subplot(5, 5, i+1)
    ax.imshow(synthetic_images[rand_idx])
    ax.set_title(class_labels[labels[rand_idx]])
    ax.axis('off')
plt.show()

313/313 [==============================] - 3s 11ms/step
Latent Vector Dim: (10000, 128)	Generated Images Dim: (10000, 32, 32, 3)

Observations

We note that the shape of the different classes can be seen clearly. However one issue is the images from other classes have generated the same images. This suggest that there might have been an issue with the generator when generating images.

improve_cond_gan_fid_class = GAN_FID(batch_size=512, noise=128,
                                     sample_size=10000, buffer_size=1024, labels=True)
improve_cond_gan_fid_class.fit(
    generator=improve_cond_gan.generator, train_data=x_train)
fid_score = improve_cond_gan_fid_class.evaluate()

Computing Real Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 5s 124ms/step
16/16 [==============================] - 2s 125ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 128ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 127ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 146ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 129ms/step
16/16 [==============================] - 2s 130ms/step
Computing Generated Image Embeddings

  0%|          | 0/20 [00:00<?, ?it/s]

16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 130ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 132ms/step
16/16 [==============================] - 2s 131ms/step
16/16 [==============================] - 2s 135ms/step
16/16 [==============================] - 2s 157ms/step
16/16 [==============================] - 3s 158ms/step
16/16 [==============================] - 3s 160ms/step
16/16 [==============================] - 3s 161ms/step
16/16 [==============================] - 3s 163ms/step
16/16 [==============================] - 3s 163ms/step
16/16 [==============================] - 3s 159ms/step
16/16 [==============================] - 3s 158ms/step
16/16 [==============================] - 3s 161ms/step
16/16 [==============================] - 3s 165ms/step
16/16 [==============================] - 3s 164ms/step
Computed Embeddings	Real Images Embedding Shape: (10240, 2048)	Generated Images Embedding Shape: (10240, 2048)
The computed FID score is: 93.59766081757371

Observations

We see the FID score is lower than the original cGAN model from 94.536161025068 to 93.59766081757371. This suggest that the changes made did improve the model. We also note that the images for certain classes are the same image which might have affected the FID score.

Saving Models

As we have the models save in the individual folders, we will also need to save it in the models folder. As we will be generating 1000 images, we will be using the 2 models that have the lowest FID score which is the ACGAN and new cGAN. We will also load the DCGAN to do a comparison.

DCGAN

gan.generator.save("models/DCGAN.h5")
gan.generator.summary()

WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "generator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_18 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_7 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_22 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_165 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 p_re_lu_19 (PReLU)          (None, 4, 4, 256)         4096      
                                                                 
 conv2d_transpose_23 (Conv2D  (None, 8, 8, 128)        524416    
 Transpose)                                                      
                                                                 
 batch_normalization_166 (Ba  (None, 8, 8, 128)        512       
 tchNormalization)                                               
                                                                 
 p_re_lu_20 (PReLU)          (None, 8, 8, 128)         8192      
                                                                 
 conv2d_transpose_24 (Conv2D  (None, 16, 16, 64)       131136    
 Transpose)                                                      
                                                                 
 batch_normalization_167 (Ba  (None, 16, 16, 64)       256       
 tchNormalization)                                               
                                                                 
 p_re_lu_21 (PReLU)          (None, 16, 16, 64)        16384     
                                                                 
 conv2d_transpose_25 (Conv2D  (None, 32, 32, 3)        3075      
 Transpose)                                                      
                                                                 
=================================================================
Total params: 3,050,691
Trainable params: 3,049,795
Non-trainable params: 896
_________________________________________________________________

ACGAN

aux_cond_gan.generator.save("models/ACGAN.h5")
aux_cond_gan.generator.summary()

WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_16 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_14 (Dense)               (None, 2048)         284672      ['concatenate_16[0][0]']         
                                                                                                  
 batch_normalization_150 (Batch  (None, 2048)        8192        ['dense_14[0][0]']               
 Normalization)                                                                                   
                                                                                                  
 leaky_re_lu_40 (LeakyReLU)     (None, 2048)         0           ['batch_normalization_150[0][0]']
                                                                                                  
 reshape_5 (Reshape)            (None, 2, 2, 512)    0           ['leaky_re_lu_40[0][0]']         
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2754752     ['reshape_5[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,050,691
Trainable params: 3,045,699
Non-trainable params: 4,992
__________________________________________________________________________________________________

New cGAN

improve_cond_gan.generator.save("models/New cGAN.h5")
improve_cond_gan.generator.summary()

WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_10 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_8 (Dense)                (None, 2048)         284672      ['concatenate_10[0][0]']         
                                                                                                  
 batch_normalization_126 (Batch  (None, 2048)        8192        ['dense_8[0][0]']                
 Normalization)                                                                                   
                                                                                                  
 p_re_lu_12 (PReLU)             (None, 2048)         2048        ['batch_normalization_126[0][0]']
                                                                                                  
 reshape_3 (Reshape)            (None, 2, 2, 512)    0           ['p_re_lu_12[0][0]']             
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape_3[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

Generating 1000 images

As part of the project objective, we will be generating 1000 images. To help display the variety, we will using a .gif to show 100 images at a time [100 per class = 1000].

Utilities function

To create a gif to show, we need to use the HTML function to display the gif using image src.

def display_gif(gif):
    return HTML('<img src="{}">'.format(gif))

DCGAN

# Loading Generator from DCGAN
gan_generator = tf.keras.models.load_model("./models/DCGAN.h5")
gan_generator.summary()

WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
Model: "generator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_18 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_7 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_22 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_165 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 p_re_lu_19 (PReLU)          (None, 4, 4, 256)         4096      
                                                                 
Model: "generator_GAN"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_18 (Dense)            (None, 2048)              264192    
                                                                 
 reshape_7 (Reshape)         (None, 2, 2, 512)         0         
                                                                 
 conv2d_transpose_22 (Conv2D  (None, 4, 4, 256)        2097408   
 Transpose)                                                      
                                                                 
 batch_normalization_165 (Ba  (None, 4, 4, 256)        1024      
 tchNormalization)                                               
                                                                 
 p_re_lu_19 (PReLU)          (None, 4, 4, 256)         4096      
                                                                 
 conv2d_transpose_23 (Conv2D  (None, 8, 8, 128)        524416    
 Transpose)                                                      
                                                                 
 batch_normalization_166 (Ba  (None, 8, 8, 128)        512       
 tchNormalization)                                               
                                                                 
 p_re_lu_20 (PReLU)          (None, 8, 8, 128)         8192      
                                                                 
 conv2d_transpose_24 (Conv2D  (None, 16, 16, 64)       131136    
 Transpose)                                                      
                                                                 
 batch_normalization_167 (Ba  (None, 16, 16, 64)       256       
 tchNormalization)                                               
                                                                 
 p_re_lu_21 (PReLU)          (None, 16, 16, 64)        16384     
                                                                 
 conv2d_transpose_25 (Conv2D  (None, 32, 32, 3)        3075      
 Transpose)                                                      
                                                                 
=================================================================
Total params: 3,050,691
Trainable params: 3,049,795
Non-trainable params: 896
_________________________________________________________________

for i in range(10):
    # creating noise
    latent_z = np.random.normal(size=(100, 128))
    
    imgs = gan_generator.predict(latent_z)
    
    fig = plt.figure(figsize=(20, 20), tight_layout=True)
    for idx, img in enumerate(imgs):
        ax = fig.add_subplot(10, 10, idx+1)
        ax.imshow((img + 1) / 2)
        ax.axis('off')
    fig.savefig(fname="./images/GAN/{}.jpeg".format(i))

4/4 [==============================] - 1s 32ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 10ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 9ms/step
4/4 [==============================] - 0s 9ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 8ms/step

images = []
gan_img_paths = glob.glob("./images/GAN/*.jpeg")
for path in gan_img_paths:
    images.append(imageio.imread(path))
imageio.mimsave('./images/GAN/GAN.gif', images, duration=1)
display_gif('./images/GAN/GAN.gif')

C:\Users\Soh Hong Yu\AppData\Local\Temp\ipykernel_14940\1410452988.py:4: DeprecationWarning: Starting with ImageIO v3 the behavior of this function will switch to that of iio.v3.imread. To keep the current behavior (and make this warning disappear) use `import imageio.v2 as imageio` or call `imageio.v2.imread` directly.
  images.append(imageio.imread(path))

ACGAN

# Loading Generator from ACGAN
aux_cond_gan_generator = tf.keras.models.load_model("./models/ACGAN.h5")
aux_cond_gan_generator.summary()

WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_16 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_14 (Dense)               (None, 2048)         284672      ['concatenate_16[0][0]']         
                                                                                                  
 batch_normalization_150 (Batch  (None, 2048)        8192        ['dense_14[0][0]']               
 Normalization)                                                                                   
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_16 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_14 (Dense)               (None, 2048)         284672      ['concatenate_16[0][0]']         
                                                                                                  
 batch_normalization_150 (Batch  (None, 2048)        8192        ['dense_14[0][0]']               
 Normalization)                                                                                   
                                                                                                  
 leaky_re_lu_40 (LeakyReLU)     (None, 2048)         0           ['batch_normalization_150[0][0]']
                                                                                                  
 reshape_5 (Reshape)            (None, 2, 2, 512)    0           ['leaky_re_lu_40[0][0]']         
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2754752     ['reshape_5[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,050,691
Trainable params: 3,045,699
Non-trainable params: 4,992
__________________________________________________________________________________________________

for i in range(10):
    # creating conditions and noise
    conditions = np.full(100, i)
    conditions = tf.keras.utils.to_categorical(conditions, 10)
    latent_z = np.random.normal(size=(100, 128))
    
    imgs = aux_cond_gan_generator.predict([latent_z, conditions])
    
    fig = plt.figure(figsize=(20, 20), tight_layout=True)
    for idx, img in enumerate(imgs):
        ax = fig.add_subplot(10, 10, idx+1)
        ax.imshow((img + 1) / 2)
        ax.axis('off')
    fig.suptitle(list(class_labels.values())[i], y=1, fontsize=28)
    fig.savefig(fname="./images/ACGAN/{}.jpeg".format(list(class_labels.values())[i]))

4/4 [==============================] - 0s 9ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 6ms/step
4/4 [==============================] - 0s 6ms/step
4/4 [==============================] - 0s 7ms/step
4/4 [==============================] - 0s 11ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 7ms/step
4/4 [==============================] - 0s 12ms/step

images = []
gan_img_paths = glob.glob("./images/ACGAN/*.jpeg")
for path in gan_img_paths:
    images.append(imageio.imread(path))
imageio.mimsave('./images/ACGAN/ACGAN.gif', images, duration=1)
display_gif('./images/ACGAN/ACGAN.gif')

C:\Users\Soh Hong Yu\AppData\Local\Temp\ipykernel_14940\899859476.py:4: DeprecationWarning: Starting with ImageIO v3 the behavior of this function will switch to that of iio.v3.imread. To keep the current behavior (and make this warning disappear) use `import imageio.v2 as imageio` or call `imageio.v2.imread` directly.
  images.append(imageio.imread(path))

New cGAN

# Loading Generator from New cGAN
cond_gan_generator = tf.keras.models.load_model("./models/New cGAN.h5")
cond_gan_generator.summary()

c:\Users\Soh Hong Yu\anaconda3\envs\gpu_env\lib\site-packages\keras\initializers\initializers_v2.py:120: UserWarning: The initializer RandomNormal is unseeded and being called multiple times, which will return identical values  each time (even if the initializer is unseeded). Please update your code to provide a seed to the initializer, or avoid using the same initalizer instance more than once.
  warnings.warn(

WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
Model: "generator_cGAN"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 Latent_Noise_Vector_z (InputLa  [(None, 128)]       0           []                               
 yer)                                                                                             
                                                                                                  
 Conditions_y (InputLayer)      [(None, 10)]         0           []                               
                                                                                                  
 concatenate_10 (Concatenate)   (None, 138)          0           ['Latent_Noise_Vector_z[0][0]',  
                                                                  'Conditions_y[0][0]']           
                                                                                                  
 dense_8 (Dense)                (None, 2048)         284672      ['concatenate_10[0][0]']         
                                                                                                  
 batch_normalization_126 (Batch  (None, 2048)        8192        ['dense_8[0][0]']                
 Normalization)                                                                                   
                                                                                                  
 p_re_lu_12 (PReLU)             (None, 2048)         2048        ['batch_normalization_126[0][0]']
                                                                                                  
 reshape_3 (Reshape)            (None, 2, 2, 512)    0           ['p_re_lu_12[0][0]']             
                                                                                                  
 Base_Generator (Sequential)    (None, 16, 16, 64)   2784320     ['reshape_3[0][0]']              
                                                                                                  
 Output_Layer (Conv2DTranspose)  (None, 32, 32, 3)   3075        ['Base_Generator[0][0]']         
                                                                                                  
==================================================================================================
Total params: 3,082,307
Trainable params: 3,076,419
Non-trainable params: 5,888
__________________________________________________________________________________________________

for i in range(10):
    # creating conditions and noise
    conditions = np.full(100, i)
    conditions = tf.keras.utils.to_categorical(conditions, 10)
    latent_z = np.random.normal(size=(100, 128))
    
    imgs = aux_cond_gan_generator.predict([latent_z, conditions])
    
    fig = plt.figure(figsize=(20, 20), tight_layout=True)
    for idx, img in enumerate(imgs):
        ax = fig.add_subplot(10, 10, idx+1)
        ax.imshow((img + 1) / 2)
        ax.axis('off')
    fig.suptitle(list(class_labels.values())[i], y=1, fontsize=28)
    fig.savefig(fname="./images/cGAN/{}.jpeg".format(list(class_labels.values())[i]))

4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 6ms/step
4/4 [==============================] - 0s 6ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 11ms/step
4/4 [==============================] - 0s 8ms/step
4/4 [==============================] - 0s 12ms/step
4/4 [==============================] - 0s 11ms/step
4/4 [==============================] - 0s 10ms/step
4/4 [==============================] - 0s 10ms/step

images = []
gan_img_paths = glob.glob("./images/cGAN/*.jpeg")
for path in gan_img_paths:
    images.append(imageio.imread(path))
imageio.mimsave('./images/cGAN/cGAN.gif', images, duration=1)
display_gif('./images/cGAN/cGAN.gif')

C:\Users\Soh Hong Yu\AppData\Local\Temp\ipykernel_14940\2200930972.py:4: DeprecationWarning: Starting with ImageIO v3 the behavior of this function will switch to that of iio.v3.imread. To keep the current behavior (and make this warning disappear) use `import imageio.v2 as imageio` or call `imageio.v2.imread` directly.
  images.append(imageio.imread(path))

Summary

We note that most of the images generated are fairly realistic, but some of the images have distorted features.

There are some issues with the images generated by the models.

1. Some of the wheels of the vehicles are warped or have no wheels present at all.
2. Some of the animals like horses and deer have thin or fat bodies.
3. We also note that some of the images are very corrupted by the background and noise.
4. Most of the animals have very weird parts combined together

We also want to note that the model has a much easier time with vehicles generation than the other classes. We also note that classes like cats, dogs and birds have learnt the texture of the animal but some of the shapes are warped. This is likely due to the variety of images and styles that the classes have and the GAN model learning and having a hard time to forms the parts of the animal.

Comparing final_model VS other people's model

Based on PapersWithCode.com, our final model is very far away from the FID score. We note that even though the ACGAN's FID score is 88.7613090961658 which is very far away from the State of the Art Image Generation on Cifar-10 [Lowest FID is 1.77]. [https://paperswithcode.com/sota/image-generation-on-cifar-10]